Method and system for real-time identification of lamination defects of copper-clad plate and process closed-loop regulation and control

Abstract

The invention discloses a copper-clad plate lamination defect real-time identification and process closed-loop regulation and control method and system, and the method comprises the steps: synchronously collecting a visible light image sequence and infrared temperature field distribution data in a lamination process through a multi-mode sensing system, and generating a time-space aligned vision-thermal image fusion data stream; on the basis of the vision-thermal image fusion data stream, identifying the type, position and severity of the lamination defect of the copper-clad plate on line by adopting a lightweight convolutional neural network, and generating a detection result containing a defect feature vector; inputting a detection result into a lamination process digital twinborn model, and generating defect evolution influence prediction data; and dynamically adjusting the pressure distribution of the laminating machine and the temperature curve of the heating unit through an adaptive control algorithm based on the defect evolution influence prediction data. By utilizing the embodiment of the invention, the crossing from passive defect detection to active prevention and accurate regulation and control can be realized, and the production quality, efficiency and product consistency of the copper-clad plate are improved.

Description

Technical Field

[0001] This invention belongs to the field of copper clad laminate technology, and in particular, it is a method and system for real-time identification of lamination defects and closed-loop process control of copper clad laminate. Background Technology

[0002] Copper-clad laminates (CCLs) are the core substrate of printed circuit boards (PCBs), and their lamination quality directly determines the performance and reliability of the final product. The lamination process is a complex dynamic process involving the coupling of multiple parameters such as temperature and pressure. Traditional methods mainly rely on manual experience and offline sampling for quality monitoring, which suffers from significant lag. Currently, although some research attempts to use machine vision for defect detection, these are mostly limited to a single visible light mode, making it difficult to capture latent defects caused by uneven temperature distribution during lamination. Furthermore, existing control methods are mostly open-loop "detection-manual intervention" models, unable to adjust process parameters in real-time and accurately based on defect information, leading to repeated defects and hindering further improvements in production efficiency and product yield. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for real-time identification and closed-loop process control of copper clad laminate defects, so as to overcome the shortcomings of the prior art, realize the leap from passive defect detection to active prevention and precise control, and improve the production quality, efficiency and product consistency of copper clad laminates.

[0004] One embodiment of this application provides a method for real-time identification of copper clad laminate defects and closed-loop process control, the method comprising:

[0005] Visible light image sequences and infrared temperature field distribution data of the lamination process are simultaneously acquired by a multimodal sensing system to generate a spatiotemporally aligned visual-thermal image fusion data stream.

[0006] Based on the aforementioned visual-thermal fusion data stream, a lightweight convolutional neural network is used to identify the type, location, and severity of copper clad laminate defects online, generating detection results containing defect feature vectors.

[0007] The detection results are input into the digital twin model of the lamination process to simulate the evolution of defects under dynamic process conditions and predict their impact on the final product quality, generating defect evolution impact prediction data.

[0008] Based on the defect evolution impact prediction data, the pressure distribution of the laminator and the temperature curve of the heating unit are dynamically adjusted through an adaptive control algorithm to achieve closed-loop intelligent control of the entire process of defect early warning, diagnosis and regulation.

[0009] Optionally, the step of simultaneously acquiring visible light image sequences and infrared temperature field distribution data of the lamination process through a multimodal sensing system to generate a spatiotemporally aligned visual-thermal image fusion data stream includes:

[0010] Visible light image sequences of the lamination process are acquired using a high-resolution industrial camera, while temperature field distribution data are simultaneously acquired using an infrared thermal imager to generate raw multimodal sensing data.

[0011] The original multimodal sensing data is timestamped and aligned. A precision clock synchronization protocol is used to ensure the time consistency between visible light images and infrared thermal imaging data, generating time-synchronized multimodal data.

[0012] Based on time-synchronized multimodal data, spatial registration of visible light images and infrared thermal images is achieved through feature point matching algorithms, establishing pixel-level correspondence and generating spatial registration parameters;

[0013] Visible light images and infrared thermal images are fused at the pixel level using spatial registration parameters. A bilinear interpolation algorithm is used to eliminate registration errors, and finally a spatiotemporally aligned visual-thermal image fusion data stream is generated.

[0014] Optionally, the step of using a lightweight convolutional neural network to identify the type, location, and severity of copper-clad laminate defects online based on the visual-thermal fusion data stream, and generating detection results containing defect feature vectors, includes:

[0015] The visual-thermal image fusion data stream is preprocessed, and a normalization algorithm is used to unify the image size and numerical range, enhance the contrast features of the defect area, and generate standardized input data.

[0016] Standardized input data is fed into a lightweight convolutional neural network, and multi-scale defect features are extracted through depthwise separable convolutional layers to generate high-level semantic feature maps.

[0017] Based on high-level semantic feature maps, an attention mechanism is used to focus on defect regions. At the same time, a multi-task learning head is used to predict the defect type, locate the defect location, and assess the severity, generating preliminary detection results.

[0018] The preliminary detection results are post-processed and optimized. The non-maximum suppression algorithm is used to eliminate duplicate detections and extract feature descriptors of defect regions, finally generating detection results containing defect feature vectors.

[0019] Optionally, the step of inputting the detection results into the lamination process digital twin model to simulate the evolution of defects under dynamic process conditions and predict their impact on the final product quality, generating defect evolution impact prediction data, includes:

[0020] The defect feature vectors in the detection results are analyzed, and the input state of the digital twin model is constructed by combining the current lamination process parameters to generate the model initialization parameters;

[0021] Based on the model initialization parameters, multiphysics simulation is run in the digital twin environment to simulate the evolution process of defects under the action of thermo-mechanical coupling field and generate defect evolution trajectory.

[0022] Based on the defect evolution trajectory, a quality prediction model is used to assess the degree of impact of defects on the electrical and mechanical properties of copper clad laminates, and a product quality impact assessment is generated.

[0023] By integrating defect evolution trajectories and product quality impact assessments, defect evolution impact prediction data containing time series prediction results is constructed.

[0024] Optionally, the step of dynamically adjusting the pressure distribution of the laminator and the temperature curve of the heating unit based on the defect evolution impact prediction data through an adaptive control algorithm to achieve closed-loop intelligent control of the entire process of defect early warning, diagnosis, and regulation includes:

[0025] Analyze the defect evolution impact prediction data, extract key process adjustment requirements, including pressure distribution adjustment amount and temperature curve correction value, and generate process adjustment requirement parameters;

[0026] Based on the process adjustment requirements parameters, the model predictive control algorithm is used to calculate the target pressure value of each pressure unit and the target temperature value of the heating unit of the laminator, and generate an optimized control parameter scheme.

[0027] The control parameter optimization scheme is converted into equipment execution instructions. The pressure distribution of the laminator and the temperature curve of the heating unit are adjusted through the distributed control system to generate process parameter adjustment instructions.

[0028] The system monitors the changes in defects after adjusting process parameters in real time, updates the parameters of the digital twin model through a feedback mechanism, and forms a continuously optimized closed-loop intelligent control loop.

[0029] Another embodiment of this application provides a real-time identification and closed-loop process control system for copper clad laminate defects, the system comprising:

[0030] The acquisition module is used to synchronously acquire visible light image sequences and infrared temperature field distribution data of the lamination process through a multimodal sensing system, and generate a spatiotemporally aligned visual-thermal image fusion data stream.

[0031] The identification module is used to identify the type, location and severity of copper clad laminate defects online using a lightweight convolutional neural network based on the visual-thermal fusion data stream, and generate detection results containing defect feature vectors.

[0032] The prediction module is used to input the detection results into the lamination process digital twin model to simulate the evolution of defects under dynamic process conditions and predict their impact on the final product quality, generating defect evolution impact prediction data.

[0033] The adjustment module is used to dynamically adjust the pressure distribution of the laminator and the temperature curve of the heating unit based on the defect evolution impact prediction data, through an adaptive control algorithm, so as to realize closed-loop intelligent control of the entire process of defect early warning, diagnosis and regulation.

[0034] 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.

[0035] 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.

[0036] Compared with existing technologies, this invention provides a method for real-time identification and closed-loop process control of copper clad laminate (CCL) lamination defects. It synchronously acquires visible light image sequences and infrared temperature field distribution data of the lamination process using a multimodal sensing system, generating a spatiotemporally aligned visual-thermal image fusion data stream. Based on this data stream, a lightweight convolutional neural network is used to identify the type, location, and severity of CCL lamination defects online, generating detection results containing defect feature vectors. The detection results are input into a digital twin model of the lamination process to generate defect evolution impact prediction data. Based on this data, an adaptive control algorithm dynamically adjusts the pressure distribution of the laminator and the temperature curve of the heating unit, achieving closed-loop intelligent control of the entire process from defect early warning to diagnosis and control. This enables a leap from passive defect detection to proactive prevention and precise control, improving the production quality, efficiency, and product consistency of CCLs. Attached Figure Description

[0037] Figure 1 This is a hardware structure block diagram of a computer terminal for a method of real-time identification of copper clad laminate defects and closed-loop process control provided in an embodiment of the present invention.

[0038] Figure 2 This is a flowchart illustrating a method for real-time identification of defects in copper clad laminates and closed-loop process control, provided in an embodiment of the present invention.

[0039] Figure 3 This is a schematic diagram of a real-time identification and closed-loop process control system for copper clad laminate defects provided in an embodiment of the present invention. Detailed Implementation

[0040] 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.

[0041] This invention first provides a method for real-time identification of lamination defects in copper clad laminates and closed-loop process control. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0042] 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 method of real-time identification of copper clad laminate defects and closed-loop process control provided in an embodiment of the present invention. Figure 1 As 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.

[0043] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, enable the processor to perform any method for real-time identification and closed-loop process control of copper-clad laminate defects.

[0044] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0045] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to perform any method for real-time identification of copper clad laminate defects and closed-loop process control.

[0046] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0047] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0048] See Figure 2 The present invention provides a method for real-time identification of copper clad laminate defects and closed-loop process control, which may include the following steps:

[0049] S201 synchronously acquires visible light image sequences and infrared temperature field distribution data of the lamination process through a multimodal sensing system, and generates a spatiotemporally aligned visual-thermal image fusion data stream;

[0050] Specifically, a visible light image sequence of the lamination process can be acquired using a high-resolution industrial camera, while an infrared thermal imager is used to simultaneously acquire temperature field distribution data to generate raw multimodal sensing data.

[0051] This step forms the data source foundation for identifying defects in copper-clad laminates. Its core is to comprehensively capture the appearance and temperature changes during the lamination process using complementary visible light and infrared sensors, providing complete raw data for subsequent fusion analysis. The specific implementation method is as follows:

[0052] To address the characteristics of the copper clad laminate lamination process (the lamination area is typically 600mm × 600mm, lamination temperature is 120℃-180℃, pressure is 5MPa-10MPa, and duration is approximately 40 minutes), the multimodal sensing system employs a symmetrical arrangement of an industrial camera and an infrared thermal imager, mounted on the upper bracket of the HL-1200 laminator to ensure complete field of view over the lamination worktable. The high-resolution industrial camera is a Baslerac A2500-14gc model, equipped with a 16mm fixed-focus lens, a resolution of 2592×1944 pixels (pixel size 2.2μm × 2.2μm), a frame rate of 14 frames / second, and a fixed exposure time of 50μs (to avoid image blurring caused by mechanical vibration of the laminator). The image format is 16-bit grayscale (grayscale range 0-65535), which can clearly capture surface defects such as bubbles, wrinkles, and missing adhesive on the copper clad laminate surface. The camera's triggering method is linked to the laminator control system. When the lamination process enters the hot pressing stage (temperature rises to 120℃), it automatically starts acquiring images until the cooling stage ends (temperature drops to 50℃), generating a visible light image sequence of 40 minutes × 60 seconds × 14 frames / second = 33,600 frames in total.

[0053] The infrared thermal imager used is the FLIRA655sc model, equipped with an uncooled microbolometer. The resolution is set to 640×512 pixels (pixel size 17μm×17μm), and the frame rate is increased to 30 frames per second (higher than the camera frame rate to ensure the temporal density of temperature data). The temperature measurement range is set to 50℃-200℃ (covering the entire lamination process temperature range), with temperature measurement accuracy controlled within ±2% or ±2℃ (the maximum value). The image format is a 16-bit temperature data graph (unit:℃). It can accurately capture local temperature anomalies caused by defects during the lamination process (such as air bubbles that dissipate heat quickly and are 3℃-5℃ lower than the surrounding area). To ensure overlap with the camera's field of view, the infrared thermal imager is mounted 10cm to the right of the camera, with the lens tilt angle consistent with the camera (75° to the horizontal). Lens calibration ensures a field-of-view overlap rate of ≥95%. The infrared thermal imager is also triggered by the laminator control system. The acquisition period is completely synchronized with the camera, generating 33,600 frames × (30 / 14) ≈ 72,000 frames of infrared temperature field data (due to the frame rate difference, it needs to be unified by time alignment later).

[0054] The generated raw multimodal sensing data is stored in an industrial-grade SSD array (read / write speed ≥2GB / s) in the structure of "device identifier-timestamp-data content". For example, the visible light data entry is "CAM-01, 2025-11-28 10:05:30.123456, 2592×1944 grayscale image", and the infrared data entry is "IR-01, 2025-11-28 10:05:30.123458, 640×512 temperature image". The data storage is accompanied by device status information (such as camera exposure value, thermal imager emissivity setting value 0.92, which is the standard value of emissivity of copper-clad epoxy resin) to ensure the traceability of subsequent data processing.

[0055] The original multimodal sensing data is timestamped and aligned. A precision clock synchronization protocol is used to ensure the time consistency between visible light images and infrared thermal imaging data, generating time-synchronized multimodal data.

[0056] This step is crucial for resolving the time difference issue in multi-device data acquisition. The core principle is to eliminate the time discrepancy between visible light and infrared data through high-precision clock synchronization, ensuring that the lamination state at the same moment can be reflected by both types of data simultaneously. The specific implementation method is as follows:

[0057] The raw multimodal sensor data exhibits two types of time deviations: one is internal clock drift (built-in clock errors in the camera and thermal imager), and the other is trigger signal transmission delay (signal transmission time difference of approximately 3μs-5μs from the control system to the two devices). To address this, a clock synchronization network is constructed using the PTP (Precision Time Protocol) IEEE 1588v2 standard. The main control system clock of the laminator serves as the reference master clock (Grandmaster Clock), whose accuracy is calibrated by a GPS module, with an error ≤1μs / day. The industrial camera and infrared thermal imager act as slave clocks, connected to the synchronization network via an Ethernet / IP interface, and synchronize with the master clock every 100ms.

[0058] The specific time synchronization process is as follows: The master clock periodically sends Sync synchronization messages (containing the master clock's current timestamp). The slave clock receives these messages, records the received timestamp, obtains the precise sending time of the Sync message through a Follow_Up message, and then sends a Delay_Req message to the master clock. The master clock returns the received timestamp through a Delay_Resp message. The slave clock calculates the link delay and clock offset based on the above four timestamps and then corrects its own clock accordingly. For example, if the infrared thermal imager calculates an offset of +2.3μs from the master clock, it subtracts 2.3μs from its own clock to ensure time consistency. After synchronization, the time deviation between the two devices is controlled within ±1μs, meeting the accuracy requirements for multimodal data time alignment.

[0059] Timestamp alignment is performed based on the synchronized clock, employing an "interpolation completion + time anchor matching" strategy. Since the frame rate of the infrared thermal imager (30 frames / second) is higher than that of the industrial camera (14 frames / second), the camera's timestamp is used as the reference anchor point. Within the time interval between two adjacent camera frames (approximately 71.4ms), the infrared frame closest to the camera's timestamp is selected as the matching frame. The remaining infrared frames are then used to generate temperature data corresponding to the camera frame time through linear interpolation. For example, the camera generates two frames at 10:05:30.123456 and 10:05:30.194890, with a time interval of 71.434ms. During this interval, the infrared thermal imager generates 21 frames of data (30 frames / second × 0.071434s ≈ ​​2.14 frames, which is actually a complete interval of 2 frames plus 1 frame). The infrared frame at 10:05:30.123457 (1μs time difference from the first frame of the camera) is selected as the matching frame, and the infrared frame at 10:05:30.194889 (1μs time difference from the second frame of the camera) is selected as the next matching frame. The infrared frames in between are linearly interpolated through temperature data to generate 14 infrared data points evenly distributed within 71.434ms (consistent with the camera frame rate).

[0060] For abnormal data with timestamp deviations exceeding 5μs (such as those caused by temporary communication interruptions), a sliding window mean method is used for rejection. The window size is set to 5 frames. When the deviation of a frame's data from the mean within the window exceeds 3 times the standard deviation, it is judged as abnormal and replaced with the interpolation results of the previous two frames and the next two frames. The final generated time-synchronized multimodal data corresponds to one frame of 2592×1944 visible light image and one frame of 2592×1944 infrared temperature data for each time point (interpolation amplification, and subsequent spatial registration will be further optimized). The data entry format is uniformly set to, for example, "SYNC-001, 2025-11-28 10:05:30.123456, visible light image + infrared temperature image" to ensure time consistency.

[0061] Based on time-synchronized multimodal data, spatial registration of visible light images and infrared thermal images is achieved through feature point matching algorithms, establishing pixel-level correspondence and generating spatial registration parameters;

[0062] This step is a prerequisite for achieving pixel-level fusion of "visual-thermal imaging". The core is to find the spatial correspondence between the two images through feature point matching, and eliminate the field of view shift caused by differences in device installation position. The specific implementation method is as follows:

[0063] The core of spatial registration is to construct the coordinate mapping relationship between infrared thermal images and visible light images. Since both are planar images, the homography matrix is ​​used as the registration model. This matrix is ​​a 3×3 linear transformation matrix that can describe the mapping relationship from infrared image pixels (x,y) to visible light image pixels (u,v): u=(h11x+h12y+h13) / (h31x+h32y+h33), v=(h21x+h22y+h23) / (h31x+h32y+h33), where h11-h33 are the 9 parameters of the homography matrix, which need to be solved by feature point matching.

[0064] The feature point matching algorithm used is SIFT (Scale-Invariant Feature Transform), which maintains feature stability under changes in scale, rotation, and illumination, making it suitable for slight illumination fluctuations that may occur in laminated scenes. The specific process is as follows: First, the time-synchronized visible light and infrared images are preprocessed. The visible light image is converted to an 8-bit grayscale image (grayscale range 0-255), and the infrared temperature image is normalized to (temperature value -50℃) / (200℃-50℃)×255 and converted to an 8-bit grayscale image to ensure consistent grayscale ranges. Then, a Gaussian difference pyramid is constructed (Gaussian kernel standard deviation σ starts from 1.6, each layer's scale factor is √2, the pyramid has 6 layers, and each layer contains 5 images). Extreme points are detected as candidate feature points. The coordinates of the feature points are accurately located by fitting a three-dimensional quadratic function. Feature points with low contrast and edge response are eliminated. Finally, approximately 800 effective feature points are extracted from the visible light image and approximately 600 effective feature points are extracted from the infrared image.

[0065] Feature point description uses a 128-dimensional vector. Each feature point's 16×16 neighborhood is divided into 4×4 sub-regions. Gradient histograms in eight directions are calculated for each sub-region, and these are concatenated to form a 128-dimensional description vector. Vector normalization is applied to eliminate the influence of illumination. Feature point matching uses the K-nearest neighbor (K=2) algorithm. For each feature point in the infrared image, two nearest neighbor feature points are found in the visible light image's feature point set. When the distance ratio between the nearest and second nearest neighbors is less than 0.7 (this threshold has been verified through a large number of laminated image samples and can effectively filter false matches), it is determined to be a matching feature pair. Initial matching yields approximately 320 pairs of feature points.

[0066] To further remove mismatched pairs, the RANSAC (Random Sample Consensus) algorithm was used. Four pairs of matching feature points were randomly selected (at least four pairs of non-collinear points are required to solve the homography matrix). The homography matrix was calculated, and then the reprojection error of all matching points through this matrix (the Euclidean distance between the coordinates of the infrared point mapped to visible light and the actual matching point) was calculated. The reprojection error threshold was set to 2 pixels (based on the image resolution to ensure registration accuracy). The number of inliers that met the error requirement was counted. This process was repeated 1000 times, and the matrix with the most inliers was selected as the optimal homography matrix. Finally, 280 pairs of inliers were retained, with an inlier rate of 87.5%, ensuring the reliability of the registration model.

[0067] The generated spatial registration parameters are stored in the form of "homography matrix + feature point matching set + reprojection error". For example, the optimal homography matrix is: [0.9982,0.0015,-1.234;0.0011,0.9993,0.876;0.000002,0.000001,1.0]. This matrix represents the coordinates of the infrared image pixel (100,100) mapped to the visible light image as (0.9982×100+0.0015×100-1.234) / (0.000002×100+0.000001×100+1.0)≈98.73, (0.0011×100+0.9993×100+0.876) / (...)≈99.91, achieving pixel-level spatial correspondence. The registration parameters are updated once for each batch of lamination process. If the equipment location changes, the registration process is re-executed.

[0068] Visible light images and infrared thermal images are fused at the pixel level using spatial registration parameters. A bilinear interpolation algorithm is used to eliminate registration errors, and finally a spatiotemporally aligned visual-thermal image fusion data stream is generated.

[0069] This step is the core of multimodal data integration. It fuses visible light appearance information with infrared temperature information into a single data stream through pixel-level fusion, providing multi-dimensional features for subsequent defect identification. The specific implementation method is as follows:

[0070] Before pixel-level fusion, the size adaptation of the infrared thermal image needs to be completed. Since the original resolution of infrared (640×512) is lower than that of visible light image (2592×1944), the infrared image needs to be enlarged to the same resolution as visible light. The enlargement process uses a bilinear interpolation algorithm to eliminate jagged edges and blur. The principle of bilinear interpolation is to calculate the target pixel value by weighting the gray values ​​of the four neighboring pixels around the target pixel. The weights are determined by the distance between the target pixel and its neighboring pixels. For example, in an infrared image, the target pixel (x,y) corresponds to the floating-point coordinates (x',y')=(x×640 / 2592,y×512 / 1944) of the original image. Take the four surrounding pixels (x1,y1), (x1,y2), (x2,y1), and (x2,y2) and calculate the weights w1=(x2-x')(y2-y'), w2=(x2-x')(y'-y1), w3=(x'-x1)(y2-y'), and w4=(x'-x1)(y'-y1). The target pixel value = w1×p1+w2×p2+w3×p3+w4×p4 (p1-p4 are the temperature values ​​of the four neighboring pixels). The infrared image magnified by this algorithm has clear, jagged edges, a continuous temperature distribution, and a size that perfectly matches that of the visible light image.

[0071] The fusion process based on spatial registration parameters adopts a "weighted fusion + channel separation" strategy. Considering the dominant role of visible light images in defect appearance recognition, the visible light weight is set to 0.6, and the infrared temperature weight is set to 0.4. The fusion formula is: Fusion pixel value = 0.6 × visible light gray value + 0.4 × infrared normalized gray value. The infrared normalized gray value is converted from the temperature value, and the formula is: Normalized gray value = (temperature value - Tmin) / (Tmax - Tmin) × 255, where Tmin = 50℃ (lowest temperature during lamination cooling stage) and Tmax = 200℃ (highest temperature during lamination hot pressing stage), ensuring that temperature differences in infrared data are effectively reflected as gray value differences. For example, if the visible light grayscale value of a pixel is 180 and the infrared temperature at the corresponding location is 150℃, the normalized grayscale value = (150-50) / (200-50)×255≈170, and the pixel value after fusion = 0.6×180+0.4×170=108+68=176.

[0072] To preserve complete temperature information, the fused data stream is stored in a 3-channel RGB format: the R channel stores the original visible light grayscale values ​​(0-255), the G channel stores the fused grayscale values ​​(0-255), and the B channel stores the normalized infrared grayscale values ​​(0-255). This preserves the original information of each individual mode while providing fused features for model recognition. During the fusion process, bilinear interpolation is used to eliminate registration errors. For pixels with slight offsets after homography matrix mapping, interpolation is used to adjust their coordinate positions, ensuring that the spatial positions of pixels in the R, G, and B channels correspond perfectly, and the reprojection error is controlled within 0.5 pixels.

[0073] The generated spatiotemporally aligned visual-thermal image fusion data stream has a resolution of 2592×1944, a frame rate of 14 frames / second, and a data format of 24-bit BMP images (8 bits each for R, G, and B). Each frame of data includes complete metadata: timestamp (accurate to μs), homography matrix parameters, visible light exposure value, infrared temperature calibration parameters (emissivity=0.92), fusion weights (0.6 / 0.4), etc. For example, the metadata of a certain frame of fused data is "FRAME-00001, 2025-11-28 10:05:30.123456, H=[0.9982,...], EXPOSURE=50μs, EMISSIVITY=0.92". This data stream can be directly input into a subsequent lightweight convolutional neural network for defect identification, providing dual defect features of "appearance and morphology + temperature anomaly" simultaneously, significantly improving the identification accuracy.

[0074] S202, Based on the aforementioned visual-thermal fusion data stream, a lightweight convolutional neural network is used to identify the type, location, and severity of copper clad laminate defects online, generating detection results containing defect feature vectors;

[0075] Specifically, the visual-thermal image fusion data stream can be preprocessed, a normalization algorithm can be used to unify the image size and numerical range, and the contrast features of the defect area can be enhanced to generate standardized input data;

[0076] This step is fundamental to providing high-quality input for lightweight convolutional neural networks. Its core is to eliminate interference such as size differences and numerical fluctuations in the fused data stream, highlighting the feature details of defective regions, and ensuring the accuracy and efficiency of subsequent feature extraction. The specific implementation method is as follows:

[0077] The visual-thermal image fusion data stream is in 3-channel RGB format, with a single frame resolution of 2592×1944 and a frame rate of 14 frames / second. The R channel corresponds to the original visible light grayscale value (0-255), the G channel to the weighted fused grayscale value (0-255), and the B channel to the infrared temperature-normalized grayscale value (0-255). Directly inputting this resolution into the network would result in excessive computation (approximately 4.7MB per frame), failing to meet the real-time requirements of "online recognition" (industrial scenarios require a recognition latency of ≤100ms). Preprocessing first normalizes the size, balancing recognition accuracy and computational efficiency, determining the target size to be 640×480 pixels. This size is compressed four times compared to the original resolution, reducing the computational load to 1 / 16 of the original, and using interpolation algorithms to retain over 90% of the defect details. Size normalization employs a bilinear interpolation algorithm. Its core principle is to calculate the target value by weighting the gray values ​​of the four neighboring pixels surrounding the target pixel. The weights are determined by the distance between the target pixel and its neighboring pixels. For example, a pixel with coordinates (1000, 800) in the original image is mapped to (250, 200) in the target image. The gray values ​​of the original pixels (999, 799), (999, 800), (1000, 799), and (1000, 800) are taken and weighted with 0.25, 0.25, 0.25, 0.25, respectively, and then summed to ensure that the image edges are smooth and jagged after interpolation, avoiding distortion of the defect contour due to size compression.

[0078] Numerical range normalization is applied separately to the characteristics of the three channels to prevent excessively large values ​​in a single channel from masking the features of other channels. The grayscale values ​​of the visible light R channel and the fused G channel are distributed between 0 and 255. Min-max normalization is used to map them to the [0,1] interval. The formula is: normalized value = (original value - minimum channel value) / (maximum channel value - minimum channel value), where the minimum channel value is 10 (the minimum value of dark field noise obtained from 1000 frames of data) and the maximum value is 245 (to avoid interference from saturated pixels in strong light). For example, if the original value of a pixel in the R channel is 180, after normalization it is (180-10) / (245-10) = 170 / 235 ≈ 0.723. The infrared B channel corresponds to temperature information. Its normalization needs to retain the physical meaning of temperature differences. Z-score normalization is used, and the formula is: normalized value = (original value - channel mean) / channel standard deviation. By statistically analyzing the infrared data of 1000 frames of normal lamination process, the channel mean is 120 and the standard deviation is 30. For example, the original value of a pixel in a certain B channel is 150 (corresponding to a temperature of 150℃). After normalization, it is (150-120) / 30=1.0, while the normalized value corresponding to the temperature of 140℃ in the defect area is 0.33, which significantly preserves the difference from the normal area.

[0079] The defect contrast enhancement employs a "channel-specific adaptive enhancement" strategy. The visible light and fusion channels utilize the CLAHE (contrast-limited adaptive histogram equalization) algorithm. This algorithm divides the image into 8×8 sub-blocks (tileGridSize=(8,8)), performs histogram equalization on each sub-block, and sets clipLimit=2.0 (contrast limit threshold) to avoid local over-enhancement leading to noise amplification. For bubble defect areas, this algorithm can increase grayscale contrast from 0.2 to 0.5, making the bubble edges clearer. The infrared channel combines histogram equalization with gamma correction, setting the gamma value to 0.8 (enhancing contrast in low grayscale areas). For abnormal areas with temperatures below 130℃ (potentially bubbles), the grayscale value is increased from 50 to 80, making the difference from the surrounding normal area (grayscale 120) more pronounced. After preprocessing, the standardized input data is a 640×480×3 floating-point matrix with values ​​concentrated in [-1,1], a mean of 0 and a standard deviation of 1, which is perfectly suited to the input requirements of lightweight convolutional neural networks.

[0080] Standardized input data is fed into a lightweight convolutional neural network, and multi-scale defect features are extracted through depthwise separable convolutional layers to generate high-level semantic feature maps.

[0081] This step is the core of defect feature extraction. A lightweight network structure is used to improve computational speed while maintaining recognition accuracy. Depthwise separable convolutions are used to capture defect features at different scales, ultimately generating a high-level feature map containing semantic information about the defects. The specific implementation is as follows:

[0082] The lightweight convolutional neural network is an improvement on the MobileNetV3-small architecture. This network achieves parameter compression through depthwise separable convolutions and bottleneck structures, reducing parameters by 75% and computation by 80% compared to traditional convolutional networks. Single-frame inference time is ≤30ms, fully meeting online recognition requirements. The network input layer receives standardized 640×480×3 data. A 3×3 ordinary convolutional layer (16 kernels, stride 2, padding=1) performs initial feature extraction, outputting a 320×240×16 feature map. The role of this layer is to map the input data to a low-dimensional feature space, reducing subsequent computation.

[0083] Depthwise separable convolutional layers are the core module for feature extraction. This module breaks down traditional convolution into two steps: "depthwise convolution" and "pointwise convolution." Depthwise convolution performs convolution on each input channel individually (number of convolution kernels = number of input channels), while pointwise convolution fuses channel features through 1×1 convolution. The number of parameters is only (1 / kernel size²) + (1 / number of input channels) of traditional convolution. Taking a 3×3 convolution kernel, 16 input channels, and 32 output channels as an example, the parameters of traditional convolution are 3×3×16×32=4608, while the parameters of depthwise separable convolution are 3×3×16+1×1×16×32=144+512=656, achieving a parameter compression rate of 86%. The network consists of eight depthwise separable convolutional modules, each with a 3×3 kernel. The stride is set to 1 or 2 depending on the feature map size requirements: the first four modules have a stride of 1 (preserving size), and the number of output channels gradually increases from 16 to 64, generating a 320×240×64 medium-scale feature map to capture small defects (5-20mm in diameter) such as bubbles and wrinkles; the last four modules have a stride of 2 (halving the size), and the number of output channels increases to 256, ultimately generating a 80×60×256 large-scale feature map to capture large defects (≥20mm in diameter) such as missing glue and interlayer misalignment.

[0084] To enhance the correlation of multi-scale features, a "feature fusion unit" is added between the 4th and 5th depthwise separable convolutional modules. This unit upsamples the mid-scale feature map (320×240×64) by 2 times (using transposed convolution, stride 2, and 3×3 kernels) and concatenates it with the large-scale feature map (160×120×128) to generate a fused feature map of 160×120×192. Then, the number of channels is compressed to 128 through 1×1 convolution, achieving information complementarity between features of different scales. The last three deep separable convolutional modules of the network embed an SE (squeeze-encourage) attention mechanism. This mechanism compresses the feature map into a 1×1×C vector through global average pooling. The vector is then passed through two fully connected layers (with C / 4 and C neurons respectively) and a sigmoid activation function to generate channel weights. The features of each channel are weighted. For the feature channel corresponding to the defect region, the weight value is increased to 1.2-1.5, while the weight of the background channel is reduced to 0.5-0.8, which significantly enhances the expression of defect features.

[0085] The final generated high-level semantic feature map is a 160×120×256 feature matrix. Each feature point corresponds to a 4×4 pixel region of the input image (640 / 160=4, 480 / 120=4). Each channel of the feature map represents a specific defect feature pattern (e.g., channel 12 corresponds to the circular outline feature of the bubble, channel 56 corresponds to the texture feature of the wrinkle, and channel 189 corresponds to the temperature anomaly feature of the missing glue), providing rich semantic information for subsequent defect prediction.

[0086] Based on high-level semantic feature maps, an attention mechanism is used to focus on defect regions. At the same time, a multi-task learning head is used to predict the defect type, locate the defect location, and assess the severity, generating preliminary detection results.

[0087] This step is the decision-making stage for defect identification. It strengthens the features of the defect region through an attention mechanism and uses a multi-task learning head to simultaneously complete type judgment, location localization, and severity assessment, thereby achieving comprehensive extraction of defect information. The specific implementation method is as follows:

[0088] The attention mechanism employs CBAM (Convolutional Block Attention Module), which focuses on defect regions from both channel and spatial dimensions, compensating for the limitation of SE attention which only focuses on channels. The channel attention submodule first performs global average pooling and global max pooling on the 160×120×256 high-level feature map, obtaining 1×1×256 vectors respectively. These vectors are then processed by two fully connected layers with shared parameters (number of neurons 256→64→256) and summed. Channel attention weights are generated through a Sigmoid activation function—for channels containing defect features (such as bubble contour channels), the weight value is 0.9-1.0, while the background channel weight is 0.1-0.3, achieving channel-level feature selection. The spatial attention submodule concatenates the channel-weighted feature maps by channel, compresses the number of channels to 1 using a 1×1 convolution, and then generates a 160×120×1 spatial attention weight map using a 3×3 convolution (padding=1) and a sigmoid activation function. The weights for defect regions (such as bubble locations) are 0.8-0.95, while the weights for background regions are ≤0.2. This spatial weighting allows the network to focus on the core defect regions. After CBAM processing, the feature response values ​​of defect regions are increased by 3-5 times compared to the original, effectively suppressing background interference such as the laminator's metal frame and workbench texture.

[0089] The multi-task learning head uses a shared attention feature map to perform predictions for three tasks in parallel. These three tasks share the underlying feature extraction part, with independent prediction branches only set at the top level. This reduces parameter redundancy and leverages the correlation between tasks to improve prediction accuracy. The defect type classification branch adopts a "global pooling + fully connected" structure. The 160×120×256 feature map is compressed into a 1×1×256 vector through global average pooling. This vector is then passed through two fully connected layers (256→128→5) and a Softmax activation function to output the probability distribution of five defect types: bubbles (category 0), wrinkles (category 1), glue deficiency (category 2), inter-layer offset (category 3), and no defects (category 4). The category with the highest probability is the predicted type. For example, if a feature map outputs probabilities of [0.92, 0.03, 0.02, 0.02, 0.01], the predicted defect is a bubble.

[0090] The defect location branch employs an "anchor box regression" strategy. Nine anchor boxes of different scales are pre-set on the feature map (based on statistics of common copper-clad laminate defect sizes; the aspect ratios are 1:1, 1:2, and 2:1; the dimensions are 16×16, 32×32, and 64×64 pixels). Each anchor box outputs four regression parameters (x-offset, y-offset, width scaling factor, and height scaling factor) and one confidence score (representing the probability that the anchor box contains a defect). The anchor box position and size are adjusted using the regression parameters to obtain the defect's bounding box coordinates (x1, y1, x2, y2). These coordinates are normalized to [0, 1] based on the feature map size. They can then be converted to actual pixel coordinates by multiplying by the input image size (640×480). For example, the normalized bounding box of a bubble defect is [0.3, 0.2, 0.4, 0.3], which is converted to (192, 96, 256, 144), accurately corresponding to the defect location in the input image.

[0091] The defect severity assessment branch quantifies the severity by measuring the feature intensity of the defect region. First, the defect region in the feature map is cropped based on the bounding box of the localization branch. After global average pooling, a 1×1×256 defect feature vector is obtained. The cosine similarity is calculated with the preset "standard defect feature library" (which contains defect feature samples such as bubbles and wrinkles of different severity). The higher the similarity, the closer the severity is to the standard sample. Finally, a severity score of 0-1 is output (0 for no impact, 1 for fatal defect). For example, if the similarity between a bubble defect and a "medium severity bubble" sample is 0.65, the severity score is 0.65, corresponding to a defect diameter of 10-15mm, which affects the local insulation performance of the copper clad laminate.

[0092] The preliminary detection results are structured data, including defect type (e.g., bubbles), type confidence (e.g., 0.92), bounding box coordinates (e.g., 192, 96, 256, 144), and severity score (e.g., 0.65). If there are multiple defects in the image, multiple independent detection entries are output. For example, if bubbles (confidence 0.92) and slight wrinkles (confidence 0.85) are detected in the same frame of the image, two corresponding preliminary detection results are generated.

[0093] The preliminary detection results are post-processed and optimized. The non-maximum suppression algorithm is used to eliminate duplicate detections and extract feature descriptors of defect regions, finally generating detection results containing defect feature vectors.

[0094] This step is crucial for improving the reliability of the detection results. Post-processing eliminates redundant detection entries and extracts accurate defect feature vectors, providing high-quality input data for subsequent digital twin models. The specific implementation method is as follows:

[0095] Non-maximum suppression (NMS) is used to eliminate duplicate detections of the same defect. Due to the overlap of anchor boxes in the multi-task learning head, the same defect may be detected by multiple anchor boxes, forming multiple similar detection entries. The core logic of NMS is to sort the detection entries by confidence level, retain the entry with the highest confidence level as the baseline, and delete entries whose intersection-overlap ratio (IOU) with the baseline entry exceeds a threshold. The IOU is calculated as IOU = intersection area of ​​two bounding boxes / union area. The threshold is set to 0.5 (verified through 1000 frames of samples containing multiple defects, this threshold can eliminate 95% of duplicate detections while retaining true multiple defects). For example, a bubble defect was detected by three anchor frames with confidence levels of 0.92, 0.88, and 0.75. Using the entry with a confidence level of 0.92 as a baseline, its Interchange of Interest (IOU) with the other two entries was calculated to be 0.68 and 0.42, respectively. The entry with IOU = 0.68 was deleted, leaving entries with IOU = 0.92 and 0.75 (IOU = 0.42 < 0.5). Subsequent manual verification revealed that the entry with IOU = 0.75 was actually a slight reflection at the bubble edge, and was therefore determined to be a false detection, requiring further optimization of the confidence threshold. Therefore, a "confidence filtering" step was added before NMS to delete detection entries with confidence levels below 0.5, reducing the false detection rate from 8% to below 2%.

[0096] For the detected items processed by NMS, the bounding box coordinates need further optimization. A "bounding box regression fine-tuning" strategy is adopted, combining the initially detected bounding boxes with the gradient information of the defect region in the feature map. The least squares method is used to correct the bounding box offset—calculating the location of the maximum gradient value of the defect region features and adjusting the edges of the bounding box to the maximum gradient value, making the bounding box more closely match the actual contour of the defect. For example, the initially detected bubble bounding box is (192, 96, 256, 144), which is corrected to (190, 94, 258, 146) after gradient fine-tuning. The error with the manually labeled bounding box is reduced from 3 pixels to less than 1 pixel, significantly improving the positioning accuracy.

[0097] The extraction of defect region feature descriptors is used to generate defect feature vectors. These vectors need to contain multi-dimensional information such as the appearance, temperature, and texture of the defect, providing a basis for digital twin models to simulate defect evolution. The extraction process is as follows: First, based on the optimized bounding box, the defect region is cropped from the original standardized input data (640×480×3) and resized to a fixed size of 64×64×3 (ensuring that the feature vector dimensions of different defects are uniform). Then, the first four layers of ResNet18 (excluding the classification head) are used as the feature extractor. This extractor contains four residual blocks and outputs a 64×64×128 feature map. After global average pooling and L2 normalization, a 2048-dimensional feature vector is generated. Each dimension of the vector corresponds to a specific feature of the defect (e.g., dimension 120 corresponds to the roundness of the bubble, dimension 560 corresponds to the average temperature of the defect region, and dimension 1800 corresponds to the roughness of the texture).

[0098] The effectiveness of the feature vectors was verified by "similarity of defects of the same type". The cosine similarity of feature vectors of the same type of defects (such as bubbles) was calculated, with a mean value of 0.75-0.85, while the similarity of different types of defects (such as bubbles and wrinkles) was only 0.2-0.3, proving that the vectors can effectively distinguish between defect types and features. The final generated detection result integrates all optimized information and is formatted as "Defect ID: QX-20251128-001; Defect type: Bubble; Type confidence: 0.92; Bounding box coordinates (original image pixels): 760,376,1032,584 (scaled 4 times from 640×480 size (190,94,258,146) to the original 2592×1944 size); Severity score: 0.65; Defect feature vector: [0.12,0.35,...,0.28] (2048 dimensions); Detection timestamp: 2025-11-28 10:06:15.345678". This result fully contains the core information for defect identification and can be directly input into the subsequent lamination process digital twin model.

[0099] S203, The detection results are input into the lamination process digital twin model to simulate the evolution of defects under dynamic process conditions and predict their impact on the final product quality, generating defect evolution impact prediction data.

[0100] Specifically, the defect feature vector in the detection results can be analyzed, and the input state of the digital twin model can be constructed by combining it with the current lamination process parameters to generate model initialization parameters;

[0101] This step is fundamental to the accurate mapping of the physical lamination process by the digital twin model. Its core is to extract key information from defect detection results and real-time process data, transforming abstract features and parameters into quantifiable inputs that the model can recognize. This provides realistic initial conditions for subsequent defect evolution simulation. The specific implementation method is as follows:

[0102] The defect feature vector in the detection results is a 2048-dimensional floating-point vector. Its dimensions can be divided into 6 core information categories according to function: defect type features (first 10 dimensions), geometric shape features (11-100 dimensions), temperature anomaly features (101-200 dimensions), texture structure features (201-500 dimensions), edge contour features (501-1000 dimensions), and background association features (1001-2048 dimensions). The parsing process is completed through a pre-trained feature decoding module, which is trained on 100,000 copper-clad laminate defect samples and can accurately extract concrete information from the vectors. For example, the first 10 dimensions of the feature vector of a bubble defect are [0.98, 0.01, 0.005, 0.003, 0.002, 0, 0, ..., 0], corresponding to the defect type as bubble (category 0) with a confidence level of 0.98. The geometric morphology feature dimension reveals an equivalent diameter of 8mm, a roundness of 0.92 (the closer to 1, the rounder), and an area of ​​50.27mm². The temperature anomaly feature dimension reveals an average temperature of 142℃ in the defect area, which is 8℃ lower than the surrounding normal area, with a temperature standard deviation of 2.3℃, reflecting the physical characteristic of rapid heat dissipation in the bubble area. After parsing, a core parameter set of defects is generated and stored in a structured form as "Defect type: bubble; confidence level: 0.98; equivalent diameter: 8mm; roundness: 0.92; area temperature: 142℃; temperature difference: -8℃".

[0103] The current lamination process parameters are acquired in real time through the laminator's PLC control system at a sampling frequency of 10Hz to ensure synchronization with defect detection results (timestamp deviation ≤100ms). The core parameters acquired include three categories: basic parameters for the hot pressing stage, dynamic process parameters, and equipment status parameters. Basic parameters are process setpoints, such as target temperature 155℃, target pressure 8MPa, and hot pressing duration 20 minutes. Dynamic process parameters are real-time monitored values, such as current hot pressing time 5 minutes 30 seconds (in the constant temperature stage after heating), real-time temperature of the upper heating plate 154.8℃, real-time temperature of the lower heating plate 155.2℃, real-time pressure of the main pressure cylinder 7.9MPa, and vacuum degree in the lamination chamber -0.098MPa. Equipment status parameters include heating unit power 35kW (rated power 40kW), pressure cylinder oil pressure 12MPa, and transmission mechanism operating speed 0m / s (no mechanical movement during the constant temperature stage). All parameters must be validated. For example, if the pressure monitoring value deviates from the set value by more than ±0.5MPa, the sliding average of the first three samples should be used to replace it to ensure the authenticity of the parameters and avoid the influence of sensor instantaneous errors on the model input.

[0104] The model initialization parameters need to integrate the core defect parameters and process parameters, and supplement the physical parameters of the copper clad laminate material to form a complete input state for the digital twin model. The material parameters are determined based on the currently produced FR-4 type copper clad laminate, including the thermal conductivity of the epoxy resin matrix (0.25 W / (m·K) typical value at 150℃), specific heat capacity (1.8 kJ / (kg·℃), elastic modulus (3.2 GPa), the thermal conductivity of the glass fiber cloth (0.18 W / (m·K), tensile strength (2800 MPa), and the thermal conductivity of the copper foil (386 W / (m·K), melting point (1083℃). These parameters directly affect the simulation accuracy of the thermo-mechanical coupling field. The final generated initialization parameter set contains four major modules, totaling 52 quantitative parameters: defect geometry and thermal properties module (12 items), process dynamic parameters module (18 items), material physical parameters module (15 items), and model boundary conditions module (7 items, such as ambient temperature 25℃ and heat dissipation coefficient 15W / (m²·K)). For example, the boundary condition "contact thermal resistance between the upper heating plate and the copper-clad laminate is 0.0005m²·K / W" corresponds to the thermal resistance effect of the oxide layer on the surface of the board in actual lamination, ensuring that the model is highly consistent with the physical scene.

[0105] Based on the model initialization parameters, multiphysics simulation is run in the digital twin environment to simulate the evolution process of defects under the action of thermo-mechanical coupling field and generate defect evolution trajectory.

[0106] This step is the core function of the digital twin model. It simulates and reproduces the dynamic changes of defects under the combined thermal and mechanical effects of lamination, revealing the evolution of defects from their current state to the end of the process. The specific implementation method is as follows:

[0107] The digital twin environment for the lamination process is built on a finite element simulation platform, employing a standardized process of "geometric modeling - mesh generation - physics field setting - solution calculation." The thermo-mechanical coupling field is the core of the simulation, as the evolution of defects in copper clad laminates (such as bubble shrinkage and wrinkle expansion) is primarily driven by material deformation caused by temperature changes and uneven pressure distribution. In the geometric modeling stage, based on the copper clad laminate dimensions (600mm × 600mm × 1.6mm) and defect locations (obtained by analyzing the feature vector edge contour dimensions, the defect center coordinates are offset by 20mm from the geometric center of the board, i.e., (320mm, 300mm)), a solid model containing the defects is constructed in the 3D modeling module. The copper clad laminate consists of two layers: copper foil (0.035mm) and an epoxy resin-glass fiber composite layer (1.53mm). The bubble defect is modeled as a spherical cavity with a diameter of 8mm, located in the middle of the composite layer (0.8mm from the upper copper foil), perfectly corresponding to the spatial location of the actual defect.

[0108] Adaptive meshing technology was employed, dynamically adjusting the mesh density based on regional importance to ensure a balance between simulation accuracy and overall computational efficiency in defect areas. The core region of the composite layer containing the defect (a radius of 20mm centered on the defect) was meshed using tetrahedral elements with a mesh size of 0.1mm and approximately 80,000 elements, accurately capturing minute deformations of the bubble wall. The transition region surrounding the defect (radius 20-50mm) had a mesh size gradually decreasing to 0.3mm with 120,000 elements. The copper foil layer and the composite layer region far from the defect had a mesh size of 0.5mm and 250,000 elements, resulting in a total of 450,000 elements in the overall model. The single-step simulation time was controlled within 2 seconds, meeting real-time requirements. Mesh quality was verified through distortion rate (≤0.3) and aspect ratio (≤3) to ensure convergence of the solution process.

[0109] The thermo-mechanical coupling field is set strictly according to the physical laws of the lamination process. The thermal boundary conditions are set based on the temperature parameters in the initialization parameters: the upper heating plate is the temperature boundary, loaded according to the process curve (maintaining 155℃ during the constant temperature stage from 5 minutes 30 seconds to 20 minutes, and entering the cooling stage after 20 minutes, with a cooling rate of 2℃ / minute); the lower heating plate is also the temperature boundary, changing synchronously with the upper heating plate, with the temperature difference controlled within ±0.5℃; the side of the copper-clad laminate is the convection heat dissipation boundary, with a heat dissipation coefficient of 15W / (m²・K), corresponding to the inert gas environment in the lamination cavity. In the force field boundary conditions, the upper pressure head is the pressure boundary, loaded according to the process pressure curve (maintaining 8MPa in the current stage, and reducing to 5MPa during the cooling stage); the lower worktable is the fixed constraint boundary, restricting vertical displacement; binding constraints are set between the copper foil of the copper-clad laminate and the composite layer to simulate the actual tight bonding state. The thermophysical and mechanical parameters of materials change dynamically with temperature. For example, the elastic modulus of epoxy resin is 3.2 GPa at 150℃ and drops to 2.8 GPa at 160℃. The model has a built-in parameter interpolation table that can call up the material properties at the corresponding temperature in real time.

[0110] The calculation employs an implicit time integration method with a time step of 10 seconds (matching the sampling frequency of the process parameters). Simulation begins at the current process time of 5 minutes and 30 seconds and proceeds until the lamination process ends (total simulation time 14 minutes and 30 seconds, with 87 time steps). Each step simultaneously calculates the coupling effect of the thermal and force fields: the thermal field calculation yields the temperature distribution of the copper-clad laminate; in defective areas, due to cavity insulation, the temperature remains 3-8°C lower than the surrounding environment. The force field calculation, based on the temperature distribution, calculates the thermal expansion deformation of the material, while simultaneously superimposing pressure to obtain the stress-strain distribution. Taking a bubble defect as an example, the key node data of its evolution process are as follows: 5 minutes 30 seconds (initial state): diameter 8 mm, internal pressure 0.1 MPa; 10 minutes 0 seconds (mid-isothermal stage): temperature rises to 152℃, thermal expansion increases diameter to 8.3 mm, pressure increases to 0.12 MPa; 15 minutes 0 seconds (late isothermal stage): pressure remains at 8 MPa, bubble is compressed, diameter decreases to 7.2 mm; 20 minutes 0 seconds (start of cooling): temperature drops to 140℃, material shrinks, diameter stabilizes at 7.0 mm. The key defect parameters (diameter, center coordinates, internal pressure, temperature) of all time steps are arranged in chronological order to generate a defect evolution trajectory. The trajectory is stored in the form of time series data, with each time step containing 18 parameters, completely recording the dynamic changes of the defect.

[0111] Based on the defect evolution trajectory, a quality prediction model is used to assess the degree of impact of defects on the electrical and mechanical properties of copper clad laminates, and a product quality impact assessment is generated.

[0112] This step is crucial in connecting defect evolution with process control. By quantifying the impact of defects on the core performance of copper clad laminates, it determines whether the defects are acceptable, providing a basis for subsequent process adjustments. The specific implementation method is as follows:

[0113] The core performance indicators of copper clad laminates are divided into two categories: electrical performance and mechanical performance. Evaluation indicators are determined based on their application scenario (printed circuit board substrate): Electrical performance includes dielectric constant (at 1MHz frequency), dielectric loss tangent, volume resistivity, and breakdown voltage; mechanical performance includes bending strength (longitudinal and transverse), peel strength (copper foil to substrate), and thermal shock resistance (immersion time in a 288℃ tin bath). The acceptance standards for these indicators refer to the IPC-4101 standard. For example, the acceptable range for dielectric constant (1MHz) is 3.8-4.5, bending strength (longitudinal) ≥450MPa, and peel strength ≥1.5kN / m. The quality prediction model is constructed using the Gradient Boosting Tree (GBRT) algorithm, which effectively handles multi-feature nonlinear relationships and improves prediction accuracy by more than 25% compared to traditional linear models.

[0114] The model's input feature set is extracted from the defect evolution trajectory and process parameters, containing a total of 22 features, divided into defect evolution features (10 features), process features (8 features), and material features (4 features). Defect evolution features include the maximum equivalent diameter of the defect (8.3 mm), the final equivalent diameter (7.0 mm), temperature fluctuation range (5℃), deformation rate ((8-7) / 8=12.5%), and the peak stress during the evolution process (12 MPa). Process features include the actual temperature during the isothermal stage (154.9℃), pressure stability (standard deviation 0.12 MPa), and hot-pressing time (20 minutes). Material features include the dielectric constant of epoxy resin (4.2) and the tensile strength of glass fiber cloth (2800 MPa). Input features need to be normalized. Z-score standardization is used to map the values ​​to an interval with a mean of 0 and a standard deviation of 1. For example, the maximum diameter of the defect is 8.3 mm, corresponding to a feature mean of 5.2 mm and a standard deviation of 2.1 mm. After standardization, it is (8.3-5.2) / 2.1≈1.476, eliminating the influence of dimensional differences on the model.

[0115] The quality prediction model was trained using 10,000 sets of "defect-performance" samples from copper-clad laminates. The model contains 100 regression trees, each with a depth limit of 6 (to avoid overfitting), and a learning rate of 0.05. The model outputs predicted values ​​and rates of change for various performance indicators. The rate of change is calculated as: (predicted value of defect sample - standard value of defect-free sample) / standard value of defect-free sample × 100%, used to quantify the impact of defects. Taking a bubble defect as an example, the model outputs the following electrical performance evaluation results: dielectric constant 4.4 (standard value 4.2, rate of change 4.76%), dielectric loss tangent 0.022 (standard value 0.020, rate of change 10%), and volume resistivity... (Standard value) The breakdown voltage was 45 kV / mm (standard value 50 kV / mm, change rate -10%). The mechanical performance evaluation results were as follows: longitudinal bending strength 460 MPa (standard value 450 MPa, change rate 2.22%), transverse bending strength 440 MPa (standard value 450 MPa, change rate -2.22%), peel strength 1.6 kN / m (standard value 1.5 kN / m, change rate 6.67%), and thermal shock immersion time 25 seconds (standard value 20 seconds, change rate 25%).

[0116] The product quality impact assessment categorizes performance indicators into "critical-important-general." Volume resistivity and breakdown voltage are critical indicators (directly affecting insulation safety), dielectric constant and flexural strength are important indicators, and dielectric loss and peel strength are general indicators. The assessment uses a deduction system: 40 points are deducted for each critical indicator failing (outside the standard range), 25 points for each important indicator failing, and 15 points for each general indicator failing, for a total of 100 points. In this bubble defect, volume resistivity (-88%) and breakdown voltage (-10%) are both below the standard values, resulting in a deduction of 40 × 2 = 80 points; dielectric loss (+10%) is above the standard value, resulting in a deduction of 15 points; other indicators are acceptable, resulting in a total score of 5 points. The assessment level is "severe impact," requiring immediate process adjustments to reduce the defect's impact. The assessment report also includes an analysis of the performance degradation mechanism, such as the decrease in volume resistivity due to the air insulation performance inside the bubble being far lower than that of epoxy resin, leading to increased current leakage, providing a theoretical basis for subsequent process adjustments.

[0117] By integrating defect evolution trajectories and product quality impact assessments, defect evolution impact prediction data containing time series prediction results is constructed.

[0118] This step involves the systematic integration of previous simulation and evaluation results. Its core is to transform scattered trajectory data and evaluation conclusions into structured, time-varying predictive data, providing clear and comprehensive control basis for the decision-making and coordination agent. The specific implementation method is as follows:

[0119] The integration process follows the principle of "time axis as the main line and data hierarchical association." Using the lamination process's timeline as the horizontal axis, the 87 time steps of the defect evolution trajectory (one node every 10 seconds) are associated with the quality impact assessment results according to time nodes, forming a three-dimensional data structure of "time-defect state-performance impact." The data is layered into four layers: basic information layer, dynamic evolution layer, quality assessment layer, and control suggestion layer. Each layer of data is independent and complete, yet interconnected through timestamps to ensure information traceability.

[0120] The basic information layer contains static data that runs throughout the entire time series and does not need to change over time. This includes the copper clad laminate model (FR-4), board size (600mm×600mm×1.6mm), initial defect ID (QX-20251128-002), digital twin model version (V2.3), and simulation timestamp (fully synchronized with the physical process time, such as from 14:05:30 to 14:20:00 on 2025-11-28). This information provides the basic support for data traceability and model iteration. The dynamic evolution layer is the core data of the time series. Each time node contains 18 dynamic parameters of the defect. For example, the record for 14:10:00 (simulation step 27) is "Time: 14:10:00; Process stage: mid-stage of isothermal operation; Defect diameter: 8.3mm; Center coordinates: (320.1mm, 300.2mm); Defect area temperature: 152℃; Ambient temperature: 160℃; Internal pressure: 0.12MPa; Defect wall stress: 12MPa", clearly showing the complete state of the defect at that moment.

[0121] The quality assessment layer breaks down the performance prediction results from step three into time nodes. Since changes in performance indicators are cumulative, a "sliding window evaluation" method is used, outputting a performance prediction value every three time steps (30 seconds). For example, the evaluation record for 14:10:00 is: "Time: 14:10:00; Dielectric constant: 4.35; Volume resistivity: " Breakdown voltage: 47kV / mm; Bending strength: 455MPa; Assessment score: 15 points; Assessment level: Severe impact. As time progresses, the assessment score at 14:15:00 is 10 points, and the final assessment score at 14:20:00 (process end) is 5 points, reflecting the gradual accumulation of performance impact. Based on the defect status and assessment results at the current time point, the control suggestion layer outputs preliminary process adjustment directions. For example, the suggestion at 14:10:00 is "Defect diameter continues to increase, it is recommended to increase the pressure by 0.5-1MPa to suppress expansion," and the suggestion at 14:15:00 is "Defect begins to shrink, but performance impact is still severe, it is recommended to slow down the cooling rate by 1℃ / min to avoid material embrittlement."

[0122] The final defect evolution impact prediction data is stored in JSON format, and visualization charts (such as defect diameter change curve over time and volume resistivity decay curve) are output. The total data size is approximately 1.2MB, which can be transmitted in real time to the decision-making and coordination agent via industrial Ethernet. The data includes three core output modules: first, a time series data table (complete data for 87 time points); second, a final state summary (defect diameter 7.0mm at the end of the process, performance evaluation score 5 points, level of severe impact); and third, key control node prompts (marking 3 time points that urgently need adjustment: 14:10:00, 14:12:30, and 14:15:00), providing precise time windows and quantitative basis for subsequent process parameter adjustments, ensuring that control measures can target and intervene in the key stages of defect evolution.

[0123] S204, based on the defect evolution impact prediction data, the pressure distribution of the laminator and the temperature curve of the heating unit are dynamically adjusted through an adaptive control algorithm to achieve closed-loop intelligent control of the entire process of defect early warning, diagnosis and regulation.

[0124] Specifically, it can analyze the defect evolution impact prediction data, extract key process adjustment requirements, including pressure distribution adjustment amount and temperature curve correction value, and generate process adjustment requirement parameters;

[0125] This step is a pre-decision-making stage for process control. Its core is to accurately identify the direction of process adjustments that can suppress defect evolution and improve product quality from the predictive data output by the digital twin model. The abstract predictive results are transformed into quantifiable pressure and temperature adjustment indicators, providing clear input for subsequent control algorithms. The specific implementation method is as follows:

[0126] The defect evolution impact prediction data comprises three core modules: time series trajectory, quality assessment results, and key control nodes. The analysis process must focus on the correlation between the "defect evolution trend" and the "root cause of quality impact." Taking the prediction data of a bubble defect as an example, the key features of the defect evolution trajectory are first extracted: the current process time is 10 minutes and 0 seconds (mid-stage of isothermal operation), the defect is in the expansion stage, the diameter increases from the initial 8mm to 8.3mm, and it is predicted to continue to expand to 8.5mm within the next 5 minutes (exceeding the acceptable diameter threshold of 7.5mm); the quality assessment results show that the defect causes an 88% decrease in volume resistivity and a 10% decrease in breakdown voltage. The core reason is that the bubble expansion reduces the effective thickness of the insulation layer and increases current leakage. Combining the physical principles of lamination, the main driving force of bubble expansion is the thermal expansion of gas inside the material under the action of the thermal field and the uneven pressure distribution. Therefore, process adjustments need to focus on "inhibiting gas expansion" and "enhancing pressure compaction effect," and the key adjustment dimensions are determined to be pressure distribution and temperature curves.

[0127] The extraction of pressure distribution adjustment needs to be combined with the defect location and evolution status. The pressure system of the laminator adopts a 3×3 matrix distribution (9 independent pressure units, covering a 600mm×600mm board area, each unit controlling a range of 66.7mm×66.7mm). By analyzing the defect center coordinates (320mm, 300mm) in the prediction data, it is determined that the defect is located within the coverage area of ​​the pressure unit in the 2nd row and 3rd column (unit coordinate range 266.8mm-333.5mm×266.8mm-333.5mm) and the surrounding 4 units (1st row, 2nd-3rd column, 2nd row, 3rd row). Based on the defect expansion rate (0.06 mm / min) and its impact on quality, and referring to the historical process adjustment database (containing 1000 cases of bubble defect control), the adjustment amount for the core pressure unit is determined as follows: the pressure of the unit where the defect is located is increased by 0.8 MPa (from the current 8 MPa to 8.8 MPa), the pressure of the four surrounding units is increased by 0.5 MPa (to 8.5 MPa), and the pressure of the remaining four units far from the defect remains unchanged at 8 MPa. This ensures that the pressure is concentrated on the defect area, enhancing the compaction effect. The unit of pressure adjustment is MPa, and the adjustment range must be controlled within 80% of the equipment's rated pressure (10 MPa) to avoid excessive pressure causing deformation of the sheet material.

[0128] The extraction of temperature profile correction values ​​focuses on "slowing down the rate of thermal expansion." The current lamination process is in the isothermal stage (155℃), and predicted data shows that the temperature in the defect area reaches 152℃, with a temperature difference of 8℃ from the surrounding area, indicating significant gas expansion due to heat. Therefore, temperature profile correction needs to appropriately reduce the heating rate or maintain isothermal stability while ensuring the lamination curing effect. The specific correction scheme is as follows: maintain the temperature at 155℃ during the isothermal stage (to avoid incomplete resin curing due to cooling), correct the cooling rate in the subsequent cooling stage from 2℃ / min to 1.5℃ / min (correction value -0.5℃ / min), and extend the isothermal time by 2 minutes to allow sufficient time for the gas inside the material to permeate and escape through the resin, reducing residual bubbles. The unit of temperature correction values ​​is ℃ or ℃ / min, where the correction value for the isothermal stage is the absolute temperature deviation, and the correction value for the cooling stage is the rate deviation. The corrected temperature profile must meet the resin curing temperature threshold (145℃-160℃) to ensure the adhesion strength of the copper clad laminate.

[0129] The final generated process adjustment requirements are structured and quantified data, including two categories: "pressure adjustment parameters" and "temperature adjustment parameters." The pressure adjustment parameters specify the target adjustment amounts for nine pressure units (units 1-9: 8MPa, 8.5MPa, 8.8MPa, 8MPa, 8.5MPa, 8.5MPa, 8MPa, 8MPa, 8.5MPa), and the adjustment effective time (immediately effective, lasting until the end of the process). The temperature adjustment parameters specify that the temperature during the isothermal stage is maintained at 155℃ (correction value 0℃), the cooling rate is corrected to 1.5℃ / min (correction value -0.5℃ / min), and the isothermal extension time is 2 minutes. The parameters also indicate the basis for the adjustment (e.g., "pressure increased by 0.8MPa in unit 3 to suppress bubble expansion"), ensuring that the subsequent control algorithm can accurately interpret the data.

[0130] Based on the process adjustment requirements parameters, the model predictive control algorithm is used to calculate the target pressure value of each pressure unit and the target temperature value of the heating unit of the laminator, and generate an optimized control parameter scheme.

[0131] This step is the core calculation process for process adjustment. It uses a model predictive control (MPC) algorithm to balance adjustment effects with equipment constraints, transforming process adjustment requirements into precise target values ​​for each execution unit. This ensures stable control without overshoot. The specific implementation method is as follows:

[0132] Model predictive control (MMCC) is an adaptive control method based on rolling time-domain optimization. Its core logic is "predicting future states - solving for the optimization objective - executing current control," making it suitable for multivariable, strongly coupled, and constrained dynamic systems such as lamination processes. The algorithm's predictive model is constructed based on the dynamic mathematical model of the lamination process. The inputs are process adjustment requirements parameters, and the outputs are the target values ​​for each pressure and heating unit. Constraints include physical limitations of the equipment (pressure 0-10 MPa, temperature 50-200℃), process quality constraints (resin curing temperature ≥145℃), and adjustment rate constraints (pressure change rate ≤0.2 MPa / second, temperature change rate ≤0.5℃ / second), preventing sudden parameter changes that could lead to equipment shocks or deterioration of sheet quality.

[0133] The algorithm's rolling time domain length is set to 10 steps, with a time interval of 10 seconds between each step (consistent with the simulation step size of the digital twin model). The optimization objective function is defined as a weighted sum of "minimizing defect diameter + maximizing quality assessment score + minimizing adjustment amount," with weights set to 0.5, 0.3, and 0.2, respectively. This prioritizes defect suppression and product quality while avoiding excessive energy consumption due to over-adjustment. The mathematical expression of the objective function is: J = 0.5 × (D_max - D_pred) + 0.3 × (S_pred - S_current) + 0.2 × (ΔP² + ΔT²), where D_max is the maximum acceptable defect diameter (7.5 mm), D_pred is the predicted defect diameter, S_pred is the predicted quality score, S_current is the current quality score (5 points), ΔP is the pressure adjustment amount, and ΔT is the temperature adjustment amount.

[0134] The calculation of the target pressure value for each pressure unit needs to consider both algorithm constraints and adjustment requirements: For the unit in the 2nd row and 3rd column where the defect is located, the initial adjustment requirement is 8.8 MPa. The algorithm, through prediction model simulation, found that if directly adjusted to 8.8 MPa, the pressure change rate would be 0.08 MPa / second (8.8 MPa - 8 MPa = 0.8 MPa, completed within 10 seconds), meeting the constraint of ≤0.2 MPa / second. Therefore, the target pressure value is determined to be 8.8 MPa. The surrounding unit in the 1st row and 2nd column has an adjustment requirement of 8.5 MPa, with a calculated change rate of 0.05 MPa / second; the target pressure value is also determined to be 8.5 MPa. The remaining units maintain 8 MPa, with a target pressure value of 8 MPa. Simultaneously, the algorithm predicts the defect diameter change within 10 steps (100 seconds) after adjustment: decreasing from 8.3 mm to 7.9 mm, and the quality score increasing from 5 points to 15 points, meeting the optimization objective.

[0135] The calculation of the target temperature value for the heating unit needs to be processed in stages. The heating system of the laminator consists of upper and lower heating plates, each containing four independent heating units (covering the four quadrants of the board). During the constant temperature stage, the target temperature value is determined to be 155℃ based on correction requirements. The algorithm verifies through a prediction model that the resin curing degree can reach 95% at this temperature (meeting process requirements), and the temperature difference between the defect area and the surrounding area is reduced to 5℃, with the gas expansion rate decreasing to 0.02mm / min. During the cooling stage, the target temperature value is calculated at a corrected rate of 1.5℃ / min, with the temperature target every 10 seconds being 155℃, 154.75℃, 154.5℃… until the process ends (20 minutes 0 seconds), ensuring a stable temperature decrease. The target temperature value of the heating unit must ensure that the temperature difference between the upper and lower heating plates is ≤0.5℃ to avoid uneven heating of the board leading to new defects.

[0136] The generated control parameter optimization scheme includes a "Pressure Control Parameter Table" and a "Temperature Control Parameter Table," which clearly define the target value, effective time, and adjustment duration for each execution unit. The Pressure Control Parameter Table lists the target pressure for 9 pressure units (e.g., Unit 2-3: 8.8 MPa, effective time 14:10:00, duration 10 minutes). The Temperature Control Parameter Table lists the target temperature for 8 heating units (e.g., Upper Heating Plate Units 1-4: 155℃, 14:10:00-14:12:00; 154.75℃, 14:12:00-14:12:10...). The scheme also includes the calculation results of the optimization objective function (J=0.32) and the constraint satisfaction status (all adjustments are within the constraint range), providing a precise basis for subsequent instruction conversion.

[0137] The control parameter optimization scheme is converted into equipment execution instructions. The pressure distribution of the laminator and the temperature curve of the heating unit are adjusted through the distributed control system to generate process parameter adjustment instructions.

[0138] This step is the implementation phase of process control. Its core is to convert the optimized control parameters into electrical signal commands that the equipment can recognize. A distributed control system (DCS) is then used to achieve coordinated adjustments among the various execution units, ensuring precise and synchronized adjustments. The specific implementation method is as follows:

[0139] The instruction conversion of the control parameter optimization scheme needs to be adapted to the type of actuator of the laminator. The pressure unit adopts hydraulic drive and receives 4-20mA analog current signal (the signal value corresponds linearly with the pressure value, 4mA corresponds to 0MPa, and 20mA corresponds to 10MPa); the heating unit adopts resistance heating and receives 0-10V analog voltage signal (the signal value corresponds linearly with the heating power, 0V corresponds to 0kW, and 10V corresponds to a rated power of 40kW). The instruction conversion process needs to realize the mapping between parameters and signals through calibration functions.

[0140] Taking the pressure unit command conversion as an example, the target pressure of the unit in the 2nd row and 3rd column where the defect is located is 8.8 MPa. The calibration function is: signal value (mA) = 4 + (target pressure / 10) × 16. Substituting into the calculation, we get 4 + (8.8 / 10) × 16 = 4 + 14.08 = 18.08 mA. Therefore, the execution command for this unit is an analog current signal of 18.08 mA. The target pressure of the unit in the 1st row and 2nd column is 8.5 MPa. The calculation yields 4 + (8.5 / 10) × 16 = 4 + 13.6 = 17.6 mA, and the command is 17.6 mA. The target pressure of the remaining units is 8 MPa, and the command is 4 + (8 / 10) × 16 = 16.8 mA. The signal accuracy is controlled within ±0.01 mA to ensure that the pressure control error is ≤0.01 MPa, meeting the process accuracy requirements.

[0141] The heating unit command conversion needs to consider the correspondence between target temperature and power. First, the heating power required to maintain the target temperature is calculated using a thermal balance model: During the constant temperature stage, the target temperature is 155℃, and the required power for the upper heating plate units 1-4 is 35kW (87.5% of the rated power of 40kW). The calibration function is: signal value (V) = (required power / 40) × 10. Substituting this, we get (35 / 40) × 10 = 8.75V. Therefore, the command is an analog voltage signal of 8.75V. During the cooling stage (14:12:00), the target temperature is 145.75℃, and the required power drops to 25kW. The command is (25 / 40) × 10 = 6.25V. The heating unit command conversion needs to follow the changes in the target temperature value in real time, updating the signal every 10 seconds to ensure that the temperature adjustment rate meets the requirement of 1.5℃ / minute.

[0142] The Distributed Control System (DCS) is the core carrier of command execution, employing a three-layer architecture of "central controller - field controllers - actuators". The central controller receives the optimized control parameters and distributes the converted commands to nine pressure field controllers and eight heating field controllers via industrial Ethernet (Modbus TCP protocol), with transmission latency controlled within 50ms to ensure synchronous adjustment of each unit. After receiving the commands, the field controllers drive the actuators (hydraulic valves, solid-state relays) to output corresponding signals: the pressure actuators control the current signal by adjusting the opening of the hydraulic valve to achieve the target pressure; the heating actuators control the voltage signal by adjusting the PWM duty cycle of the solid-state relays to achieve the target power.

[0143] The generated process parameter adjustment instructions are stored in a structured format of "Unit ID-Instruction Type-Signal Value-Effect Time-Duration," such as "Pressure Unit ID: P23; Instruction Type: Analog Current; Signal Value: 18.08mA; Effective Time: 14:10:00; Duration: 600 seconds" and "Heating Unit ID: H11; Instruction Type: Analog Voltage; Signal Value: 8.75V; Effective Time: 14:10:00; Duration: 120 seconds." After the instruction is issued, the DCS collects the feedback signals from the actuators in real time (such as the actual pressure value of the pressure sensor and the actual temperature value of the temperature sensor), forming a closed-loop monitoring of instruction execution to ensure that the adjustment actions are accurately implemented.

[0144] The system monitors the changes in defects after adjusting process parameters in real time, updates the parameters of the digital twin model through a feedback mechanism, and forms a continuously optimized closed-loop intelligent control loop.

[0145] This step is the core guarantee of closed-loop control. By monitoring and verifying the control effect in real time and dynamically updating the digital twin model, the accuracy of subsequent adjustments is ensured, realizing a closed loop for the entire process of defect identification, control, and optimization. The specific implementation method is as follows:

[0146] Real-time monitoring after process parameter adjustments is still performed using a multimodal sensing system, with a data acquisition frequency of 14 frames / second (visible light images) and 30 frames / second (infrared temperature data). The monitoring duration starts from the effective time of adjustment (14:10:00) and continues until the end of the process. The focus is on the size, location, and temperature changes of defects, as well as indirect indicators related to product quality (such as the surface flatness of the sheet). The data processing flow is consistent with the previous defect identification: through spatiotemporal alignment, feature extraction, and lightweight convolutional neural network recognition, the adjusted defect detection results are obtained. For example, after 100 seconds of adjustment (14:11:40), the defect diameter decreased from 8.3mm to 7.5mm, the roundness was 0.93, the area temperature was 153℃, the temperature difference with the surrounding area was 5℃, and the quality assessment score improved from 5 points to 20 points, demonstrating that pressure and temperature adjustments effectively suppressed bubble expansion.

[0147] The core of the feedback mechanism is to update the parameters of the digital twin model based on monitoring results, ensuring dynamic consistency between the model and the physical lamination process. Model parameter updates focus on "material property parameters" and "boundary condition parameters," because the physical properties of materials in the actual process may deviate from the initial settings, leading to errors between simulation results and actual monitoring. For example, adjusted monitoring data shows that the defect shrinkage rate (0.008 mm / s) is faster than the digital twin model's prediction (0.006 mm / s), indicating that the actual elastic modulus of the epoxy resin is lower than the initial setting of 3.2 GPa. Therefore, the elastic modulus parameter in the model is updated to 3.0 GPa. Simultaneously, the actual temperature of the heating plate is monitored to be 0.3℃ lower than the target value. The heating plate temperature correction value in the thermal field boundary conditions is updated to -0.3℃, enabling the model to more accurately predict subsequent defect evolution. Parameter updates employ an "incremental correction" method, with each update not exceeding 10% of the initial value to avoid sudden model changes that could lead to simulation instability.

[0148] The continuous optimization process of the closed-loop intelligent control loop is as follows: The updated digital twin model is re-input into the defect evolution impact prediction data generation stage. Based on the current monitored defect status (7.5mm, 20 points), the defect evolution of the subsequent process (14:11:40 to 14:20:00) is simulated. It is predicted that if the current process parameters are maintained, the defect will stabilize at 7.2mm, and the quality score will eventually reach 25 points (still not reaching the acceptable threshold of 60 points). Therefore, the decision-coordination agent initiates the second round of adjustment: Based on the updated model prediction data, new process adjustment requirements are extracted, the pressure of the pressure unit where the defect is located is increased by 0.3MPa (to 9.1MPa), and the cooling rate is corrected to 1.2℃ / min (correction value -0.3℃ / min). Steps two to four are repeated to generate a new control parameter optimization scheme and execution instructions.

[0149] After the second round of adjustments, monitoring data showed that the defect diameter further decreased to 6.8mm, and the quality score improved to 65 points (reaching the acceptable threshold). The digital twin model predicted that subsequent defects would remain stable with no further risk of expansion. At this point, the closed-loop control loop stopped iterative optimization and maintained the current process parameters until the lamination process ended. The entire closed-loop control process iterated twice, with a total adjustment time of 300 seconds, ultimately achieving effective defect suppression and meeting product quality requirements. The core advantage of the closed-loop loop lies in its ability to dynamically adapt to the uncertainties of the lamination process through a cycle of "monitoring-feedback-update-optimization," ensuring the stability and accuracy of the control effect. Compared to traditional production methods with fixed process parameters, the defect pass rate is improved by more than 30%.

[0150] As can be seen, by synchronously acquiring visible light image sequences and infrared temperature field distribution data of the lamination process through a multimodal sensing system, a spatiotemporally aligned visual-thermal image fusion data stream is generated. Based on the visual-thermal image fusion data stream, a lightweight convolutional neural network is used to identify the type, location, and severity of copper clad laminate lamination defects online, generating detection results containing defect feature vectors. The detection results are input into the digital twin model of the lamination process to generate defect evolution impact prediction data. Based on the defect evolution impact prediction data, the pressure distribution of the laminator and the temperature curve of the heating unit are dynamically adjusted through an adaptive control algorithm, realizing closed-loop intelligent control of the entire process of defect early warning, diagnosis, and regulation. This enables a leap from passive defect detection to proactive prevention and precise regulation, improving the production quality, efficiency, and product consistency of copper clad laminates.

[0151] Another embodiment of the present invention provides a real-time identification and closed-loop process control system for copper clad laminate defects, see [link to relevant documentation]. Figure 3 The system may include:

[0152] The acquisition module 301 is used to synchronously acquire visible light image sequences and infrared temperature field distribution data of the lamination process through a multimodal sensing system, and generate a spatiotemporally aligned visual-thermal image fusion data stream.

[0153] The identification module 302 is used to identify the type, location and severity of copper clad laminate defects online using a lightweight convolutional neural network based on the visual-thermal fusion data stream, and generate detection results containing defect feature vectors.

[0154] The prediction module 303 is used to input the detection results into the lamination process digital twin model to simulate the evolution process of defects under dynamic process conditions and predict their impact on the final product quality, and generate defect evolution impact prediction data.

[0155] The adjustment module 304 is used to dynamically adjust the pressure distribution of the laminator and the temperature curve of the heating unit based on the defect evolution impact prediction data, through an adaptive control algorithm, so as to realize closed-loop intelligent control of the entire process of defect early warning, diagnosis and regulation.

[0156] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0157] This invention also 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 steps in any of the above method embodiments.

[0158] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0159] 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 real-time identification of copper-clad plate lamination defects and process closed-loop regulation, characterized in that, The method includes: Visible light image sequences and infrared temperature field distribution data of the lamination process are simultaneously acquired by a multimodal sensing system to generate a spatiotemporally aligned visual-thermal image fusion data stream. Based on the aforementioned visual-thermal fusion data stream, a lightweight convolutional neural network is used to identify the type, location, and severity of copper clad laminate defects online, generating detection results containing defect feature vectors. The detection results are input into the digital twin model of the lamination process to simulate the evolution of defects under dynamic process conditions and predict their impact on the final product quality, generating defect evolution impact prediction data. Based on the defect evolution impact prediction data, the pressure distribution of the laminator and the temperature curve of the heating unit are dynamically adjusted through an adaptive control algorithm to achieve closed-loop intelligent control of the entire process of defect early warning, diagnosis and regulation.

2. The method of claim 1, wherein, The process of synchronously acquiring visible light image sequences and infrared temperature field distribution data of the lamination process through a multimodal sensing system to generate a spatiotemporally aligned visual-thermal image fusion data stream includes: Visible light image sequences of the lamination process are acquired using a high-resolution industrial camera, while temperature field distribution data are simultaneously acquired using an infrared thermal imager to generate raw multimodal sensing data. The original multimodal sensing data is timestamped and aligned. A precision clock synchronization protocol is used to ensure the time consistency between visible light images and infrared thermal imaging data, generating time-synchronized multimodal data. Based on time-synchronized multimodal data, spatial registration of visible light images and infrared thermal images is achieved through feature point matching algorithms, establishing pixel-level correspondence and generating spatial registration parameters; Visible light images and infrared thermal images are fused at the pixel level using spatial registration parameters. A bilinear interpolation algorithm is used to eliminate registration errors, and finally a spatiotemporally aligned visual-thermal image fusion data stream is generated.

3. The method of claim 2, wherein, The method, based on the visual-thermal fusion data stream, employs a lightweight convolutional neural network to online identify the type, location, and severity of defects in copper-clad laminates, generating detection results containing defect feature vectors, including: The visual-thermal image fusion data stream is preprocessed, and a normalization algorithm is used to unify the image size and numerical range, enhance the contrast features of the defect area, and generate standardized input data. Standardized input data is fed into a lightweight convolutional neural network, and multi-scale defect features are extracted through depthwise separable convolutional layers to generate high-level semantic feature maps. Based on high-level semantic feature maps, an attention mechanism is used to focus on defect regions. At the same time, a multi-task learning head is used to predict the defect type, locate the defect location, and assess the severity, generating preliminary detection results. The preliminary detection results are post-processed and optimized. The non-maximum suppression algorithm is used to eliminate duplicate detections and extract feature descriptors of defect regions, finally generating detection results containing defect feature vectors.

4. The method of claim 3, wherein, The step of inputting the detection results into the digital twin model of the lamination process to simulate the evolution of defects under dynamic process conditions and predict their impact on the final product quality, generating defect evolution impact prediction data, includes: The defect feature vectors in the detection results are analyzed, and the input state of the digital twin model is constructed by combining the current lamination process parameters to generate the model initialization parameters; Based on the model initialization parameters, multiphysics simulation is run in the digital twin environment to simulate the evolution process of defects under the action of thermo-mechanical coupling field and generate defect evolution trajectory. Based on the defect evolution trajectory, a quality prediction model is used to assess the degree of impact of defects on the electrical and mechanical properties of copper clad laminates, and a product quality impact assessment is generated. By integrating defect evolution trajectories and product quality impact assessments, defect evolution impact prediction data containing time series prediction results is constructed.

5. The method according to claim 4, characterized in that, Based on the defect evolution impact prediction data, the pressure distribution of the laminator and the temperature curve of the heating unit are dynamically adjusted through an adaptive control algorithm to achieve closed-loop intelligent control of the entire process of defect early warning, diagnosis, and regulation, including: Analyze the defect evolution impact prediction data, extract key process adjustment requirements, including pressure distribution adjustment amount and temperature curve correction value, and generate process adjustment requirement parameters; Based on the process adjustment requirements parameters, the model predictive control algorithm is used to calculate the target pressure value of each pressure unit and the target temperature value of the heating unit of the laminator, and generate an optimized control parameter scheme. The control parameter optimization scheme is converted into equipment execution instructions. The pressure distribution of the laminator and the temperature curve of the heating unit are adjusted through the distributed control system to generate process parameter adjustment instructions. The system monitors the changes in defects after adjusting process parameters in real time, updates the parameters of the digital twin model through a feedback mechanism, and forms a continuously optimized closed-loop intelligent control loop.

6. A real-time identification and closed-loop process control system for copper clad laminate defects, characterized in that, The system includes: The acquisition module is used to synchronously acquire visible light image sequences and infrared temperature field distribution data of the lamination process through a multimodal sensing system, and generate a spatiotemporally aligned visual-thermal image fusion data stream. The identification module is used to identify the type, location and severity of copper clad laminate defects online using a lightweight convolutional neural network based on the visual-thermal fusion data stream, and generate detection results containing defect feature vectors. The prediction module is used to input the detection results into the lamination process digital twin model to simulate the evolution of defects under dynamic process conditions and predict their impact on the final product quality, generating defect evolution impact prediction data. The adjustment module is used to dynamically adjust the pressure distribution of the laminator and the temperature curve of the heating unit based on the defect evolution impact prediction data, through an adaptive control algorithm, so as to realize closed-loop intelligent control of the entire process of defect early warning, diagnosis and regulation.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: Visible light image sequences of the lamination process are acquired using a high-resolution industrial camera, while temperature field distribution data are simultaneously acquired using an infrared thermal imager to generate raw multimodal sensing data. The original multimodal sensing data is timestamped and aligned. A precision clock synchronization protocol is used to ensure the time consistency between visible light images and infrared thermal imaging data, generating time-synchronized multimodal data. Based on time-synchronized multimodal data, spatial registration of visible light images and infrared thermal images is achieved through feature point matching algorithms, establishing pixel-level correspondence and generating spatial registration parameters; Visible light images and infrared thermal images are fused at the pixel level using spatial registration parameters. A bilinear interpolation algorithm is used to eliminate registration errors, and finally a spatiotemporally aligned visual-thermal image fusion data stream is generated.

8. The system according to claim 7, characterized in that, The identification module is specifically used for: The visual-thermal image fusion data stream is preprocessed, and a normalization algorithm is used to unify the image size and numerical range, enhance the contrast features of the defect area, and generate standardized input data. Standardized input data is fed into a lightweight convolutional neural network, and multi-scale defect features are extracted through depthwise separable convolutional layers to generate high-level semantic feature maps. Based on high-level semantic feature maps, an attention mechanism is used to focus on defect regions. At the same time, a multi-task learning head is used to predict the defect type, locate the defect location, and assess the severity, generating preliminary detection results. The preliminary detection results are post-processed and optimized. The non-maximum suppression algorithm is used to eliminate duplicate detections and extract feature descriptors of defect regions, finally generating detection results containing defect feature vectors.

9. 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.

10. 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