HDI communication board defect detection method and system combined with deep learning

Abstract

This invention provides a method and system for defect detection in HDI communication boards that incorporates deep learning. Relating to the field of deep learning technology, the method first simultaneously acquires multi-layer structural images of the HDI communication board, including surface circuitry, internal vias, and interlayer dielectrics. Next, it constructs a deep learning defect detection model integrating feature extraction, co-occurrence correlation, temporal processing, and network topology generation modules. Then, the multi-layer structural images are input into the deep learning defect detection model. Feature extraction and co-occurrence correlation generate cross-layer feature co-occurrences, followed by temporal processing to reconstruct the defect evolution path and form the network topology. Finally, the network topology generation module parses the topology, locates the initial defect level and diffusion path, and generates a detection report. This invention achieves accurate and comprehensive detection of defects in HDI communication boards and can promptly discover hidden cross-layer defects.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for detecting defects in HDI communication boards that incorporates deep learning. Background Technology

[0002] In the electronics manufacturing industry, HDI (High-Density Interconnect) communication boards are widely used in various high-end electronic devices due to their advantages such as high integration and miniaturization. However, the complex structural characteristics of HDI communication boards make them prone to defects during the manufacturing process. These defects can seriously affect the performance and reliability of the communication board, and consequently affect the normal operation of the entire electronic device.

[0003] Currently, defect detection for HDI communication boards mainly relies on traditional manual visual inspection and detection methods based on simple image processing algorithms. Manual visual inspection is not only inefficient but also easily influenced by the subjective factors of the inspectors, making it difficult to guarantee the consistency and accuracy of the results. While detection methods based on simple image processing algorithms improve efficiency to some extent, they can only analyze a single level of the communication board's surface image and cannot delve into the intrinsic relationships between the board's multiple layers. This makes it difficult to accurately detect and locate some highly concealed defects that evolve across layers. For example, tiny defects in internal vias may gradually spread to other layers during the production process, and traditional inspection methods struggle to detect this cross-layer defect evolution. Therefore, existing HDI communication board defect detection methods have significant limitations and cannot meet the demands of the modern electronics manufacturing industry for high-quality, high-efficiency inspection. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for detecting defects in HDI communication boards using deep learning, the method comprising: Synchronously acquire multi-layer structure images of the HDI communication board, including surface circuit images, internal via images, and interlayer dielectric images; A deep learning defect detection model is constructed, which integrates a feature extraction module, a co-occurrence association module, a temporal processing module, and a network topology generation module. The multi-layer structured image is input into the feature extraction module of the deep learning defect detection model. The defect-related features of each layer of the image are extracted through convolution operation within the feature extraction module. The correlation between defect-related features is calculated by the co-occurrence correlation module to generate a cross-layer feature co-occurrence containing cross-layer feature correlation information. The time-series processing module of the deep learning defect detection model processes cross-layer feature co-occurrences, reconstructs the defect evolution path, and forms a defect trajectory reconstruction network topology. The defect trajectory reconstruction network topology presents the evolution process of the defect from the initial state to the manifest state. The network topology generation module of the deep learning defect detection model analyzes the defect trajectory to reconstruct the network topology, locates the initial generation level and diffusion path information of the defect, and generates a detection report containing cross-layer feature co-occurrence association and defect evolution information.

[0005] In another aspect, embodiments of the present invention also provide an HDI communication board defect detection system incorporating deep learning, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0006] Based on the above, this embodiment of the invention synchronously acquires multi-layer structure images of HDI communication boards, covering key information such as surface circuits, internal vias, and interlayer media. The constructed deep learning defect detection model integrates multiple modules, including feature extraction, co-occurrence correlation, temporal processing, and network topology generation. These modules work collaboratively to achieve a complete process from defect feature extraction to defect evolution path reconstruction. The feature extraction module accurately extracts defect-related features from each layer of the image through convolution operations. The co-occurrence correlation module further calculates the correlation between these features, generating cross-layer feature co-occurrences, fully exploring the intrinsic connections between multi-layer structures, and is able to discover cross-layer defects that are difficult to detect using traditional methods. The timing processing module processes cross-layer feature symbiosis, reconstructs the defect evolution path, and forms a defect trajectory reconstruction network topology, presenting the evolution process of the defect from its initial state to its manifest state. The network topology generation module analyzes the defect trajectory reconstruction network topology, locates the initial generation level and diffusion path information of the defect, and generates a detection report containing cross-layer feature symbiosis associations and defect evolution information. This enables accurate and comprehensive detection of defects in HDI communication boards, timely discovery of hidden defects evolving across layers, effectively improving the accuracy and reliability of detection, and ensuring the quality of HDI communication boards and the stable operation of electronic equipment. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the HDI communication board defect detection method combined with deep learning provided in the embodiments of the present invention.

[0008] Figure 2 This is a schematic diagram of the hardware architecture of the HDI communication board defect detection system combining deep learning provided in an embodiment of the present invention. Detailed Implementation

[0009] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for detecting defects in HDI communication boards that incorporates deep learning, provided in one embodiment of the present invention. The following is a detailed description of this method for detecting defects in HDI communication boards that incorporates deep learning.

[0010] Step S110: Synchronously acquire multi-layer structure images of the HDI communication board, the multi-layer structure images including surface circuit images, internal via images and interlayer dielectric images.

[0011] In this embodiment, defect detection of a certain type of HDI communication board is taken as an example. This HDI communication board has an eight-layer circuit structure. A composite acquisition system consisting of an optical imaging unit, an X-ray detection unit, and an ultrasonic scanning unit is used to achieve synchronous acquisition of multi-layer structure images through a system-level trigger signal. Surface circuit images are acquired by a surface array charge-coupled device camera with an image resolution of A pixels × B pixels, an image format of 16-bit grayscale TIFF, and a storage path in the RAID array cache directory specified by the system. Internal via images are acquired using a microfocus X-ray source in conjunction with a flat panel detector. The X-ray source tube voltage is set to C kilovolts, the tube current is set to D milliamps, and the image size is consistent with the surface circuit images. Interlayer dielectric images are acquired by a high-frequency ultrasonic scanning microscope with a scanning frequency of E megahertz. The image data is stored in DICOM standard format and includes physical property information such as dielectric thickness distribution and acoustic impedance changes. During the data acquisition process, the equipment control module fixes the HDI communication board to the vacuum adsorption stage through a precision ball screw transmission system and performs positioning calibration through a laser interferometer to ensure that the spatial coordinate system of the three acquisitions is consistent and the positioning error is controlled within the range of F micrometers, where F is three times the minimum resolution of the equipment positioning system.

[0012] Step S120: Construct a deep learning defect detection model, which integrates a feature extraction module, a co-occurrence association module, a temporal processing module, and a network topology generation module.

[0013] To address the multi-layered structure of the HDI communication board, a deep learning defect detection model was constructed using a modular, serial architecture. Each functional module interacts with the others via a standardized internal data bus. The model is developed using the PyTorch deep learning framework, employing a mixed-precision training mode. The runtime environment is a server configured with G graphics processors, each with at least H gigabytes of dedicated video memory, and a system memory configuration of I gigabytes. The model input consists of three types of image data acquired in step S110, and the output is a structured detection report containing defect location coordinates, defect type encoding, defect severity level, and defect evolution time series.

[0014] Step S121: Construct the overall architecture of the deep learning defect detection model. The overall architecture adopts a modular serial structure, which includes an input interface layer, a feature extraction module, a co-occurrence association module, a temporal processing module, a network topology generation module, and an output interface layer in sequence.

[0015] The input interface layer contains three parallel data receiving channels, corresponding to surface line images, internal via images, and interlayer dielectric images, respectively. Each channel includes a format parsing subunit, a coordinate calibration subunit, and a data buffer subunit. The format parsing subunit supports TIFF and DICOM image decoding and extracts pixel data by calling open-source image libraries. The coordinate calibration subunit, based on the device coordinate information in the image metadata, unifies the image coordinates of different acquisition devices to the physical coordinate system of the HDI communication board through affine transformation. The data buffer subunit adopts a circular queue data structure with a queue length of J to ensure continuous flow of image data. The feature extraction module receives the preprocessed image data from the input interface layer and extracts defect-related features through multi-size convolution operations. The co-occurrence association module performs association analysis on the extracted multi-layer features to generate cross-layer feature co-occurrences. The time-series processing module receives the cross-layer feature co-occurrences arranged in a time sequence and reconstructs the defect evolution path through a recurrent neural network structure. The network topology generation module converts the evolution path into network topology data and extracts key defect information. The output interface layer encapsulates the detection results into a JSON format report and provides a RESTful API interface with the production execution system to support real-time uploading of detection results.

[0016] Step S122: Set up a multi-size convolutional kernel group in the feature extraction module for extracting features of multi-layer structure images. The multi-size convolutional kernel group contains convolutional kernels of specific sizes corresponding to surface line images, internal via images and interlayer medium images, respectively. The size of each specific size convolutional kernel is set based on the pixel distribution statistics of the features in the corresponding layer image.

[0017] The feature extraction module uses an improved ResNet architecture as its basic framework, specifically the ResNet-K architecture, where K represents the number of network layers. Dedicated convolutional kernel groups are designed for different image feature levels: three kernel sizes are configured for surface line images (L×L, M×M, and N×N); two kernel sizes are configured for internal via images (O×O and P×P); and two kernel sizes are configured for interlayer medium images (Q×Q and R×R). Each kernel group is bound to its corresponding image type via a parameter configuration file, stored in XML format, containing parameters such as kernel size, number, and stride. During feature extraction, the input interface layer automatically calls the matching convolutional kernel parameters based on the image type and loads them into the feature extraction module's computation graph via dynamic graph calculation.

[0018] Step S1221: Analyze the pixel distribution of the line features in the surface line image, determine the average pixel spacing of the surface line features in the image, and preliminarily set the convolution kernel size range corresponding to the surface line feature extraction based on the average pixel spacing.

[0019] Pixel distribution analysis was performed on S surface line image samples from the training sample set. Line edges were extracted using the Canny edge detection algorithm, and the line direction was identified using Hough transform. The pixel distance between adjacent line edges was calculated. The line spacing data of all samples were statistically analyzed to calculate the average pixel spacing T. Based on this average pixel spacing T, the convolutional kernel size for surface line feature extraction was initially set to a range of [T / U]×[T / U] to [T×V]×[T×V], where U and V are empirical coefficients, with U ranging from 2 to 3 and V ranging from 4 to 6, ensuring that the convolutional kernel size can cover the width of a single line and the interval area between adjacent lines.

[0020] Step S1222: Fine-tune the initially set convolution kernel size using surface line image samples in the training sample set of the deep learning defect detection model. Apply convolution kernels of different sizes to feature extraction of surface line image samples, compare the completeness of the extracted line features, and select the convolution kernel size that can completely cover the line features.

[0021] W surface line image samples are selected from the training sample set. These samples contain surface line images with different line distribution densities and line widths. The initially set range of convolutional kernel sizes is divided into X consecutive size gradients, with a stride of Y pixels for each size gradient. For the convolutional kernel size parameters of each size gradient, a temporary feature extraction unit is constructed, with the number of network layers, activation function, pooling method, and other parameters of the temporary feature extraction unit remaining fixed. The selected surface line image samples are input one by one into each temporary feature extraction unit. The surface line features in the surface line image samples are extracted through convolutional operations within the temporary feature extraction unit, resulting in line feature extraction results for each surface line image sample under different convolutional kernel sizes. The integrity of the line contour in each line feature extraction result is evaluated by calculating the intersection-union ratio (IUU) of the extracted contour and the manually labeled contour. The line distribution coverage is evaluated by calculating whether the line distribution in each line feature extraction result completely corresponds to the actual line distribution in the surface line image samples, specifically by calculating the ratio of the number of extracted lines to the actual number of lines. Based on the combined results of the line contour integrity calculation and the line distribution coverage calculation, the line feature extraction results at each convolutional kernel size are quantitatively scored. The scoring formula is: Score = (Intersection over Union (IoU) × Z + Coverage Ratio × (1-Z)) × 100, where Z is a weighting coefficient ranging from 0.6 to 0.8. The average score of all surface line image samples at each convolutional kernel size is calculated, and the average scores of different convolutional kernel sizes are compared to select the convolutional kernel size range with the highest average score. Within the selected convolutional kernel size range, the size gradient is further subdivided with a subdivision step size of Y / 2 pixels. The above steps of surface line image sample extraction, line feature integrity evaluation, and scoring statistics are repeated to obtain the optimal convolutional kernel size. The optimal convolutional kernel size is determined as the final convolutional kernel size corresponding to surface line feature extraction and is entered into the parameter configuration file of the feature extraction module as a fixed operation parameter for surface line feature extraction.

[0022] Step S1223: Analyze the pixel distribution of the via features in the internal via image, determine the average pixel diameter of the via features in the image, and preliminarily set the convolution kernel size range corresponding to the internal via feature extraction based on the average pixel diameter.

[0023] Pixel distribution analysis was performed on AA images of internal through-holes in the training sample set. Through-hole regions were identified using circular Hough transform, and the number of pixels corresponding to the circumcircle diameter of each through-hole region was calculated. The through-hole diameter data of all samples were statistically analyzed to calculate the average pixel diameter BB. Based on this average pixel diameter BB, the convolution kernel size for internal through-hole feature extraction was initially set to a range from [BB / CC]×[BB / CC] to [BB×DD]×[BB×DD], where CC and DD are empirical coefficients, with CC ranging from 2 to 4 and DD ranging from 1.5 to 3, ensuring that the convolution kernel size can cover the through-hole region and a certain range of surrounding background area.

[0024] Step S1224: Fine-tune the initially set convolution kernel size using internal via image samples in the training sample set of the deep learning defect detection model. Apply convolution kernels of different sizes to feature extraction of internal via image samples, compare the completeness of the extracted via features, and select the convolution kernel size that can completely cover the via features.

[0025] Select EE internal via image samples from the training sample set. These samples contain internal via images with different aperture sizes and via densities. Divide the initially set convolution kernel size range into FF consecutive size gradients, with a stride of GG pixels for each size gradient. For each size gradient, construct a temporary feature extraction unit, keeping other parameters of the temporary feature extraction unit fixed. Input the selected internal via image samples one by one into each temporary feature extraction unit, and extract internal via features through convolution operations to obtain via feature extraction results under different convolution kernel sizes. Evaluate the extraction results through via contour integrity calculation and via distribution coverage calculation, and select the optimal convolution kernel size after quantification and scoring. The specific process refers to the method for selecting the convolution kernel size of surface line images in step S1222.

[0026] Step S1225: Analyze the pixel distribution of medium features in the interlayer medium image, determine the average pixel thickness of the medium features in the image, and preliminarily set the convolution kernel size range corresponding to the extraction of interlayer medium features based on the average pixel thickness.

[0027] Pixel distribution analysis was performed on the HH interlayer medium image samples in the training sample set. The medium layer region was extracted using a threshold segmentation algorithm, and the number of pixels in the thickness direction of the medium layer was calculated. The medium thickness data of all samples were statistically analyzed to calculate the average pixel thickness II. Based on this average pixel thickness II, the convolution kernel size range corresponding to the interlayer medium feature extraction was initially set to [II / JJ]×[II / JJ] to [II×KK]×[II×KK], where JJ and KK are empirical coefficients, with JJ ranging from 3 to 5 and KK ranging from 2 to 4, ensuring that the convolution kernel size can cover the thickness and lateral distribution features of the medium layer.

[0028] Step S1226: Fine-tune the initially set convolution kernel size using interlayer medium image samples in the training sample set of the deep learning defect detection model. Apply convolution kernels of different sizes to feature extraction of interlayer medium image samples, compare the completeness of the extracted medium features, and select the convolution kernel size that can completely cover the medium features.

[0029] LL interlayer medium image samples are selected from the training sample set. These samples contain interlayer medium images with different medium thicknesses and distributions. The initially set range of convolutional kernel sizes is divided into MM consecutive size gradients, with a stride of NN pixels for each size gradient. For the convolutional kernel size parameters of each size gradient, a temporary feature extraction unit is constructed, while other parameters of the temporary feature extraction unit remain fixed. The selected interlayer medium image samples are input one by one into each temporary feature extraction unit, and interlayer medium features are extracted through convolution operations to obtain medium feature extraction results under different convolutional kernel sizes. The extraction results are quantitatively scored by evaluating the uniformity of medium distribution and the accuracy of medium thickness, and the optimal convolutional kernel size is selected. The specific process is as described in step S1222.

[0030] Step S1227: Integrate the selected surface line feature extraction convolution kernel, internal via feature extraction convolution kernel, and interlayer medium feature extraction convolution kernel into a multi-size convolution kernel group for the feature extraction module. Each group of convolution kernels corresponds to a unique hierarchical image feature extraction task.

[0031] The optimal convolutional kernel sizes selected in steps S1222, S1224, and S1226 are used as the convolutional kernels for surface circuit feature extraction, internal via feature extraction, and interlayer medium feature extraction, respectively. A unique feature extraction task identifier is assigned to each group of convolutional kernels: "FL-" for surface circuit feature extraction, "FV-" for internal via feature extraction, and "FM-" for interlayer medium feature extraction. These convolutional kernel parameters are integrated into the multi-size convolutional kernel group configuration file of the feature extraction module. The configuration file uses JSON format and includes information such as convolutional kernel size, number, task identifier, and applicable image type.

[0032] Step S1228: Set the stride parameter of the convolution kernel. The stride parameter of the surface line feature extraction convolution kernel is set to a preset ratio of the average pixel spacing of the line feature. The stride parameter of the internal via feature extraction convolution kernel is set to a preset ratio of the average pixel diameter of the via feature. The stride parameter of the interlayer medium feature extraction convolution kernel is set to a preset ratio of the average pixel thickness of the medium feature.

[0033] The stride parameter of the convolutional kernel for surface line feature extraction is set to 00% of the average pixel spacing T, where 00 ranges from 30% to 50%; the stride parameter of the convolutional kernel for internal via feature extraction is set to PP% of the average pixel diameter BB, where PP ranges from 20% to 40%; and the stride parameter of the convolutional kernel for interlayer medium feature extraction is set to QQ% of the average pixel thickness II, where QQ ranges from 25% to 45%. The stride parameter settings ensure that the receptive fields of adjacent convolutional operations have a certain overlap during feature extraction, avoiding the loss of feature information.

[0034] Step S1229: Set the number of convolution kernels. Determine the number of convolution kernels according to the number of feature dimensions of each layer of the image, so that each feature dimension corresponds to a unique convolution kernel for extraction operation. A one-to-one correspondence is achieved through parameter binding.

[0035] Principal component analysis (PCA) was used to analyze the feature dimensions of each layer of the image, determining that the number of feature dimensions for the surface circuit image is RR, the number of feature dimensions for the internal via image is SS, and the number of feature dimensions for the interlayer dielectric image is TT. Accordingly, the number of convolutional kernels for surface circuit feature extraction was set to RR, the number of convolutional kernels for internal via feature extraction was set to SS, and the number of convolutional kernels for interlayer dielectric feature extraction was set to TT. Each convolutional kernel is associated one-to-one with a specific feature dimension through a parameter binding table, which is stored in the parameter database of the feature extraction module and indexed during the feature extraction process using the feature dimension identifier.

[0036] Step S12210: Input the determined convolution kernel size, stride, and number parameters into the parameter configuration file of the feature extraction module. Apply the configuration parameters to the feature extraction process through parameter loading operation in the feature extraction module to achieve the corresponding adaptation of the convolution kernel to the feature distribution density of each layer of the image.

[0037] The kernel size determined in steps S1222, S1224, and S1226, the stride parameter determined in step S1228, and the quantity parameter determined in step S1229 are entered into the parameter configuration file of the feature extraction module. The parameter configuration file is in XML format and includes elements such as kernel group labels, image type labels, size labels, stride labels, and quantity labels. During the initialization phase, the feature extraction module reads the configuration file through the parameter loading subunit, parses the parameters into tensor data structures, and loads them into the weight matrix of the convolutional layer. During feature extraction, the corresponding kernel parameters are automatically matched according to the type of the input image, achieving dynamic adaptation between the convolution kernel and the feature distribution density of each layer.

[0038] Step S123: Set up a feature preprocessing submodule in the feature extraction module. The feature preprocessing submodule performs pixel normalization operation on the input multi-layer structure image and linearly transforms the pixel values ​​of each image to a preset numerical range.

[0039] The feature preprocessing submodule includes a pixel value scaling unit and a noise filtering unit. The pixel value scaling unit uses a min-max normalization algorithm to linearly transform the pixel values ​​of the input image from the original dynamic range to the [0, 1] interval. Specifically, for each pixel in the image, its pixel value is subtracted from the minimum pixel value, and then divided by the difference between the maximum and minimum pixel values. The noise filtering unit uses a Gaussian filtering algorithm with a kernel size of UU×UU and a standard deviation of VV to suppress noise in the normalized image. The preprocessed image data is transmitted to the convolutional layer of the feature extraction module through the output interface of the feature preprocessing submodule.

[0040] Step S124: A feature cross-operation unit is set in the symbiotic association module to perform channel-by-channel feature comparison operation on the input defect-related features of different levels. The channel-by-channel feature comparison operation calculates the correlation between features of different levels by analyzing the synchronous change trend of feature response values.

[0041] The feature cross-operation unit comprises a channel alignment subunit, a response value extraction subunit, and a trend analysis subunit. The channel alignment subunit unifies the channel dimensions of the input surface line features, internal via features, and interlayer dielectric features to the same number, achieving dimension matching through zero-padding or channel clipping. The response value extraction subunit extracts the feature response value of each channel from the aligned feature map, forming a response value sequence. The trend analysis subunit employs a sliding window technique, with a window size of WW and a step size of XX, to perform sliding analysis on the response value sequence, calculating the similarity of response value change trends between feature channels at different levels. Similarity calculation uses a dynamic time warping algorithm, measuring trend similarity by calculating the minimum cumulative distance between two sequences; a smaller cumulative distance indicates a more synchronized change trend.

[0042] Step S125: Set up a weight allocation submodule in the symbiotic association module. The weight allocation submodule assigns weights to each feature component based on the correlation strength value between feature components.

[0043] The weight allocation submodule includes a correlation strength calculation unit, a weight level division unit, and a weight adjustment unit. The correlation strength calculation unit uses the Pearson correlation coefficient algorithm to calculate the correlation strength between feature components. Specifically, it standardizes the response value sequences of two feature components, calculates the covariance, and then divides it by the product of the standard deviations of the two sequences to obtain the correlation coefficient as the correlation strength value. The weight level division unit divides the correlation strength values ​​into Y levels, each corresponding to a fixed weight parameter. The higher the correlation strength value, the larger the corresponding weight parameter. The weight adjustment unit normalizes the initial weight parameters so that the sum of the weight parameters of all feature components is 1. This is achieved by dividing each weight parameter by the sum of all weight parameters.

[0044] Step S126: A loop operation unit is set in the time series processing module to receive the feature sequence arranged in the order of timestamps, and to establish the association between the features before and after in the feature sequence through the internal state transmission of the loop operation unit.

[0045] The recurrent computation unit employs a Long Short-Term Memory (LSTM) network architecture, containing ZZ hidden layers, each with AAA neurons. The network's input dimension is the same as the feature dimension of the feature sequence, while the output dimension remains consistent with the input dimension. The LTM network's gating units include an input gate, a forget gate, and an output gate. The activation function for each gating unit is the sigmoid function, and the cell state update uses the tanh function. When processing the feature sequence, the input features at each time step interact with the hidden state from the previous time step, updating the cell and hidden states through the gating mechanism to achieve the correlation modeling of preceding and following features.

[0046] Step S127: Set up a trajectory reconstruction submodule in the time series processing module. The trajectory reconstruction submodule performs interpolation operation on the serialized feature data and generates intermediate feature states based on the changing gradient between adjacent features to reconstruct a continuous defect evolution path.

[0047] The trajectory reconstruction submodule includes a gradient calculation unit, an interpolation point generation unit, and a feature state generation unit. The gradient calculation unit calculates the gradient changes of each feature component between two adjacent temporal features. This is achieved by subtracting the value of the previous temporal feature component from the value of the subsequent temporal feature component, and then dividing by the difference between the two timestamps. The interpolation point generation unit determines the number of interpolation points based on the time interval; the larger the time interval, the more interpolation points are needed. This is specifically obtained by multiplying the time interval by a preset interpolation density coefficient (BBB). The feature state generation unit uses a linear interpolation algorithm to calculate the component values ​​of intermediate feature states based on the gradient changes and the interpolation point positions, generating a continuous defect evolution path.

[0048] Step S128: Set up a network topology drawing unit in the network topology generation module to perform coordinate mapping calculations, map each evolution stage in the defect evolution path to spatial coordinate points, and connect the spatial coordinate points in chronological order to generate network topology data.

[0049] The network topology drawing unit comprises a coordinate mapping subunit and a path connection subunit. The coordinate mapping subunit maps the feature state vector of each evolutionary stage to coordinate points on a two-dimensional plane, using principal component analysis to reduce the dimensionality of the high-dimensional feature vectors to two-dimensional space, obtaining the x and y coordinate values ​​of the coordinate points. The path connection subunit connects the coordinate points with directed line segments in chronological order. The thickness of the line segments is determined by the intensity of feature changes in the corresponding evolutionary stage; the greater the intensity of change, the thicker the line segment. The generated network topology data is stored using an adjacency list data structure, containing a node list and an edge list. The node list stores the coordinate information and feature states of each evolutionary stage, while the edge list stores the connection relationships and the intensity of change.

[0050] Step S129: Set up an information extraction submodule in the network topology generation module. The information extraction submodule traverses the nodes in the network topology data, identifies the level corresponding to the node with the earliest timestamp as the initial generation level of the defect, and calculates the diffusion path parameters based on the spatial coordinate sequence and timestamp sequence of the node.

[0051] The information extraction submodule includes a node traversal unit, an initial level identification unit, and a path parameter calculation unit. The node traversal unit traverses all nodes in the network topology data in timestamp order, recording the timestamp and corresponding level information for each node. The initial level identification unit selects the node with the earliest timestamp and determines the level corresponding to that node as the initial defect generation level. The path parameter calculation unit calculates the length, direction, and velocity parameters of the diffusion path based on the spatial coordinate sequence and timestamp sequence of the nodes. The length is obtained by accumulating the Euclidean distances between adjacent nodes, the direction is calculated by the coordinate differences between adjacent nodes, and the velocity is obtained by dividing the distance by the time interval.

[0052] Step S1210: The input interface layer, feature extraction module, co-occurrence association module, temporal processing module, network topology generation module and output interface layer are connected sequentially through the internal data bus of the deep learning defect detection model to form a deep learning defect detection model.

[0053] The internal data bus employs a shared memory-based communication method, with a data transfer rate no less than CCC megabytes per second. The bus protocol includes a data packet header, data length, data type, data body, and checksum field. The output of the input interface layer is connected to the input of the feature extraction module via the data bus. The output of the feature extraction module is connected to the input of the symbiotic association module. The output of the symbiotic association module is connected to the input of the timing processing module. The output of the timing processing module is connected to the input of the network topology generation module. The output of the network topology generation module is connected to the input of the output interface layer. Data transfer between modules is scheduled through a bus controller to ensure that data is transmitted sequentially and without loss.

[0054] Step S130: Input the multi-layer structure image into the feature extraction module of the deep learning defect detection model, extract the defect-related features of each layer image through convolution operation in the feature extraction module, calculate the correlation between defect-related features through the co-occurrence correlation module, and generate a cross-layer feature co-occurrence containing cross-layer feature correlation information.

[0055] In this embodiment, the surface circuit image, internal via image, and interlayer dielectric image acquired in step S110 are input into the input interface layer of the deep learning defect detection model according to a preset category order. After format conversion and coordinate calibration of the image data, the input interface layer transmits the data to the feature extraction module. The feature extraction module performs convolution operations on the images at each layer using a multi-size convolution kernel group to extract defect-related features. The co-occurrence correlation module performs cross-comparison and weight allocation on the extracted multi-layer features to generate a cross-layer feature co-occurrence.

[0056] Step S131: The multilayer structure image is imported into the deep learning defect detection model through the input interface layer of the deep learning defect detection model in the order of surface line image, internal via image, and interlayer medium image. The image data is converted into a feature input format that can be recognized by the deep learning defect detection model through format conversion operation. The converted surface line image is input into the surface line feature extraction subunit of the feature extraction module. Through continuous sliding operation of multi-size convolution kernels, the edge contour features and line distribution features in the surface line image are extracted region by region. The surface line features are obtained after multiple rounds of convolution superposition operation.

[0057] The input interface layer's format conversion operation transforms TIFF and DICOM format image data into a model-recognizable four-dimensional tensor format, with dimensions in the order of [batch size, number of channels, height, width]. The surface line feature extraction subunit contains DDD convolutional layers, each configured with the kernel size and stride parameters selected in step S1222. The convolution operation uses a continuous sliding method, starting from the top left corner of the image and proceeding from left to right and top to bottom, with each slide using a preset stride parameter. Edge contour features are extracted using the Sobel operator convolution, and line distribution features are extracted using multi-scale convolution kernel stacking. After DDD rounds of convolution stacking operations, a surface line feature tensor with dimensions [1, EEE, FFF, GGG] is obtained, where EEE is the number of feature channels, and FFF and GGG are the height and width of the feature map.

[0058] Step S132: Input the converted internal via image into the internal via feature extraction subunit of the feature extraction module. Through multi-directional scanning convolution operation, cover all angles of the via region to extract the aperture morphology features and via connectivity features in the internal via image. After dimensional integration operation, obtain the internal via features.

[0059] The internal via feature extraction subunit contains HHH convolutional layers, each configured with the kernel size and stride parameters selected in step S1224. Multi-directional scanning convolution operations employ eight convolutional kernels, corresponding to 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° directions, respectively. Multi-angle coverage of the via region is achieved through response value fusion of the directional convolutional kernels. Aperture morphology features are extracted through circularity calculation and area statistics, while via connectivity features are extracted through region labeling and connected component analysis. Dimension integration operations utilize a combination of global average pooling and fully connected layers to integrate multi-scale features into an internal via feature tensor of dimensions [1, III, JJJ, KKK].

[0060] Step S133: Input the converted interlayer medium image into the interlayer medium feature extraction subunit of the feature extraction module. Through the adaptation operation of the stride adjustable convolution kernel, extract the bonding density features and medium distribution features in the interlayer medium image. After feature aggregation operation, obtain the interlayer medium features.

[0061] The interlayer medium feature extraction subunit contains LLL convolutional layers. The stride of the convolutional kernel is dynamically adjusted according to the distribution density of the medium features; the stride decreases in high-density regions and increases in low-density regions. Fitting density features are extracted using the energy and entropy values ​​of the gray-level co-occurrence matrix, while medium distribution features are extracted using clustering analysis with a Gaussian mixture model. Feature aggregation employs an attention mechanism, weighting and fusing feature maps from different convolutional layers. The weights are automatically learned from the response intensity of the feature maps. The aggregated interlayer medium feature tensor has dimensions [1, MMM, NNN, OOO].

[0062] Step S134: The surface circuit features, internal via features, and interlayer dielectric features are synchronously transmitted to the feature cross-operation unit of the symbiotic association module through the internal data bus of the deep learning defect detection model. The surface circuit features and internal via features are fused channel by channel. The feature response values ​​of each channel are compared, and the feature components with synchronously changing response values ​​in the two types of features are retained to obtain the circuit via association features. The circuit via association features and interlayer dielectric features are fused across channels to integrate the feature components with synchronous changing trends to obtain the cross-layer association features.

[0063] The internal data bus uses DMA transmission to synchronously transmit surface trace features, internal via features, and interlayer dielectric features to the co-occurrence correlation module. The feature cross-operation unit first performs channel-by-channel fusion of the surface trace features and internal via features. For each channel pair, it calculates the correlation coefficient of the response values ​​and retains feature components with correlation coefficients greater than a preset threshold PPP to form trace-via correlation features. Then, the trace-via correlation features and interlayer dielectric features are fused across channels, using the same correlation coefficient filtering method to retain feature components with synchronous changing trends, resulting in cross-layer correlation features. The dimensions of the cross-layer correlation features are [1, QQQ, RRR, SSS].

[0064] Step S135: The weight allocation submodule of the symbiotic association module performs weight adjustment operation on the cross-layer association features. The weight adjustment operation assigns corresponding weight parameters to each feature component based on the association strength value between features, thereby obtaining weighted cross-layer association features.

[0065] The weight allocation submodule first extracts all feature components from the cross-layer correlation features and calculates the correlation strength between each feature component and other components. Weight levels are assigned based on the correlation strength, and initial weight parameters are assigned to each feature component. These initial weight parameters are then normalized so that their sum equals 1. Next, the weight verification unit verifies and fine-tunes the weight parameters to ensure that the weighted feature components accurately reflect the defect correlation relationships. The final weighted cross-layer correlation feature has the same dimensions as the cross-layer correlation features, but each feature component is multiplied by its corresponding weight parameter.

[0066] Step S1351: The weight allocation submodule of the symbiotic association module extracts all feature components in the cross-layer association features through feature parsing operations, and determines the source level and attribute information of each feature component.

[0067] Feature parsing is performed by traversing the channel dimensions of cross-layer associated features, with each channel corresponding to a feature component. For each feature component, its source layer (surface circuit layer, internal via layer, or interlayer dielectric layer) and attribute information (such as edge contour, aperture morphology, etc.) are determined by querying the parameter configuration file of the feature extraction module. The parsing results are stored in the feature component information table, which includes fields such as component ID, source layer identifier, attribute type, and channel index.

[0068] Step S1352: Determine the correlation strength between any two characteristic components by calculating the Pearson correlation coefficient or covariance of the response values ​​between the characteristic components.

[0069] For any two feature components, their response values ​​at all spatial locations are extracted, forming two response value sequences. The Pearson correlation coefficient algorithm is used to calculate the correlation strength. Specifically, the two sequences are standardized, the covariance is calculated, and then divided by the product of the standard deviations of the two sequences. If the length of the two sequences is TTT, the covariance is calculated by the average of the product of the deviations of the response values ​​at each location, and the standard deviation is calculated by the square root of the average of the squared deviations. The correlation coefficient ranges from -1 to 1, with a larger absolute value indicating a stronger correlation.

[0070] Step S1353: Quantify the correlation strength of each pair of feature components to obtain the feature correlation strength value. The magnitude of the feature correlation strength value directly reflects the tightness of the symbiotic relationship between feature components.

[0071] The absolute value of the Pearson correlation coefficient is used as the numerical value of the feature association strength, with a value ranging from [0, 1]. A correlation coefficient of 1 indicates a perfect positive correlation, a correlation coefficient of -1 indicates a perfect negative correlation, and a correlation coefficient of 0 indicates no correlation. By taking the absolute value, negative correlations are converted into positive association strength values, thus uniformly measuring the degree of symbiotic relationship between feature components.

[0072] Step S1354: Set weight allocation rules, divide multiple weight levels according to the distribution range of feature association strength values, and each weight level corresponds to a fixed weight parameter. The higher the association strength value, the larger the corresponding weight parameter.

[0073] The feature association strength values ​​are divided into U weight levels, for example, 5 levels: [0, 0.2) corresponds to weight level 1, [0.2, 0.4) corresponds to weight level 2, [0.4, 0.6) corresponds to weight level 3, [0.6, 0.8) corresponds to weight level 4, and [0.8, 1] corresponds to weight level 5. Each weight level corresponds to a fixed weight parameter, such as 0.1 for weight level 1, 0.3 for weight level 2, 0.5 for weight level 3, 0.7 for weight level 4, and 0.9 for weight level 5. The weight parameter is set according to the principle that the higher the association strength value, the larger the weight parameter.

[0074] Step S1355: Based on the feature association strength value, match the corresponding weight level for each feature component, and then determine the initial weight parameter of each feature component. Perform normalization processing on the initial weight parameter to adjust the weight parameters of all feature components to the 0-1 range, so that the sum of the weight parameters of each feature component is 1.

[0075] For each feature component, the correlation strength values ​​between it and all other feature components are calculated, and the average value is taken as the comprehensive correlation strength of that component. Based on the comprehensive correlation strength value, the corresponding weight level is matched to obtain the initial weight parameters. Normalization is achieved by dividing each initial weight parameter by the sum of all weight parameters, ensuring that the sum of the normalized weight parameters is 1, and that the value of each weight parameter ranges from 0 to 1.

[0076] Step S1356: Construct a weight verification unit in the weight allocation submodule. The weight verification unit performs the following operations: input the features of multi-layer structured image samples with initial weight parameters into the co-occurrence association module, calculate the weighted cross-layer association features of the samples and the association relationship calculation results based on them; compare the association relationship calculation results with the real defect association relationships of the pre-labeled samples, and calculate the matching degree between the two.

[0077] The weighted verification unit selects VVV multi-layer structured image samples from the training sample set. These samples contain pre-labeled real defect associations. The sample features are input into the co-occurrence association module, which calculates weighted cross-layer association features using initial weight parameters. Then, the feature cross-operation unit mines the associations. The mined associations are compared with the pre-labeled real associations, and the number of successfully matched associations is counted. The matching degree is calculated by dividing the number of successfully matched associations by the total number of real associations.

[0078] Step S1356-1: Select multi-layer structure image samples from the training sample set of the deep learning defect detection model. The multi-layer structure image samples contain annotation information, which records whether there is a predefined correlation between surface circuit features, internal via features and interlayer medium features, and the type identifier of the correlation.

[0079] The training sample set contains WWW images of multilayer structures. Each image comprises three parts: a surface circuit image, an internal via image, and an interlayer dielectric image. The annotation information is stored in XML format and includes fields such as feature component ID, association type (e.g., causal relationship, adjoint relationship), and association strength reference value. Association type identifiers use numerical encoding, such as 1 for causal relationship, 2 for adjoint relationship, and 3 for suppression relationship.

[0080] Step S1356-2: Input the selected multilayer structure image samples into the feature extraction module of the deep learning defect detection model according to the categories of surface circuit image, internal via image, and interlayer medium image. Extract the surface circuit features, internal via features, and interlayer medium features of the multilayer structure image samples through convolution operation in the feature extraction module.

[0081] The sample image is imported into the feature extraction module through the input interface layer. After feature preprocessing and convolution operation, surface line features, internal via features, and interlayer dielectric features are extracted respectively. The feature extraction process is the same as steps S131-S133, and the extracted feature tensor dimensions are [1, EEE, FFF, GGG], [1, III, JJJ, KKK], and [1, MMM, NNN, OOO].

[0082] Step S1356-3: The extracted multi-layer image sample features are transmitted to the co-occurrence association module. The initial weight parameters are applied to perform weight adjustment operations on the multi-layer image sample features to obtain the weighted cross-layer association features of the multi-layer image samples.

[0083] The extracted sample features are transmitted to the co-occurrence association module via the internal data bus. The weight allocation submodule applies the initial weight parameters to weight the feature components, resulting in weighted cross-layer association features. The weighting process multiplies each feature component by its corresponding initial weight parameter, while the feature dimension remains unchanged after weighting.

[0084] Step S1356-4: Perform co-occurrence relationship mining operations on the weighted cross-layer association features of multi-layer structured image samples through the feature cross-operation unit of the co-occurrence association module to obtain the mining association relationship of multi-layer structured image sample features.

[0085] The feature cross-operation unit performs channel-by-channel comparisons on weighted cross-layer associated features, calculates the association strength values ​​between feature components, and filters out significantly associated feature component pairs based on preset thresholds to form mined association relationships. The mined association relationships include information such as feature component ID pairs, association strength values, and association type predictions.

[0086] Step S1356-5: Compare the mining association relationships of the features of the multi-layer structure image samples with the real defect association relationships of the multi-layer structure image samples one by one through the association relationship matching operation, and count the number of successfully matched association relationships.

[0087] The association matching operation performs a dual comparison using feature component ID pairs and association type identifiers. If the feature component ID pairs and association type identifiers of the discovered association are completely consistent with the actual association, the match is considered successful. The total number of successfully matched associations in all samples is then counted.

[0088] Step S1356-6: Calculate the ratio of the number of successfully matched associations to the total number of actual defect associations to obtain the matching degree value. Compare the calculated matching degree value with the matching degree threshold. If the matching degree value is higher than or equal to the matching degree threshold, the initial weight parameter verification is deemed successful. If the matching degree value is lower than the matching degree threshold, unmatched associations are marked. Analyze the feature components corresponding to the unmatched associations to determine the weight parameters that caused the matching failure.

[0089] The matching score is calculated by dividing the number of successful matches by the total number of actual associations, with the matching score threshold set to XXX. If the matching score is >= XXX, the initial weight parameters are validated; otherwise, unmatched associations are marked, and the weight parameters of the feature components corresponding to these associations are analyzed to determine whether the weight parameters are too low or too high, causing the associations to be not correctly discovered.

[0090] Step S1356-7: Feed back the verification results and the unmatched association analysis report to the weight allocation submodule of the symbiotic association module.

[0091] The verification results include a matching score, a pass / fail indicator, and a list of unmatched relationships. The unmatched relationship analysis report includes information such as the feature component ID, relationship type, current weight parameters, and suggested adjustment direction for each unmatched relationship. This information is fed back to the weight allocation submodule via an internal message queue for fine-tuning of the weight parameters.

[0092] Step S1357: Adjust the initial weight parameters according to the verification calculation results, fine-tune the weight parameters of the feature components whose matching degree does not meet the standard, bind the adjusted weight parameters with the corresponding feature components to form a set of feature components with weight labels, and integrate the set of feature components with weight labels into a feature vector through the weight allocation submodule of the symbiotic association module to obtain the weighted cross-layer association feature. The weight ratio of each component in the weighted cross-layer association feature corresponds to the tightness of the symbiotic relationship between features.

[0093] For weight parameters that fail validation, fine-tuning is performed based on the suggestions in the unmatched correlation analysis report, such as increasing or decreasing the weight parameters of specific feature components. The fine-tuning step size is set to YYY, and validation is performed again after each adjustment until the matching degree reaches the threshold. The adjusted weight parameters and feature components are bound to each other through a parameter binding table, forming a set of feature components with weight labels. The weighted fusion operation concatenates the weighted feature components along the channel dimension to form a feature vector, obtaining the weighted cross-layer correlation features.

[0094] Step S136: The feature integration operation unit of the symbiotic association module performs dimension unification operation on the weighted cross-layer association features, performs association solidification operation on the features after unification, and arranges the scattered symbiotic relationships in an orderly manner according to the weight ratio of the feature components to form a cross-layer feature symbiosis containing symbiotic association information of surface circuit features, internal via features and interlayer medium features.

[0095] The feature integration unit first unifies the dimensions of the weighted cross-layer correlated features by adjusting the height and width of the feature map to a fixed size of ZZZ×ZZZ using adaptive pooling. The correlation solidification operation sorts the feature components according to their weight percentages from highest to lowest. The weight percentage is calculated by multiplying the weight parameter by the feature response value. The sorted feature components are arranged sequentially to form a cross-layer feature co-occurrence body with dimensions [1, AAA×ZZZ×ZZZ], where AAA represents the number of feature channels. The cross-layer feature co-occurrence body contains the co-occurrence correlation information of features from each layer, with feature components having higher weight percentages positioned at the beginning.

[0096] Step S140: The cross-layer feature co-occurrence is processed by the temporal processing module of the deep learning defect detection model to reconstruct the defect evolution path and form a defect trajectory reconstruction network topology. The defect trajectory reconstruction network topology presents the evolution process of the defect from the initial state to the manifest state.

[0097] In this embodiment, the time-series processing module receives cross-layer feature co-occurrences arranged in a time sequence, extracts time-series change features through a cyclic operation unit, and constructs feature association chains. After path filtering and trajectory reconstruction, the defect evolution process is converted into network topology data, forming a defect trajectory reconstruction network topology.

[0098] Step S141: Using the temporal processing module of the deep learning defect detection model, the cross-layer feature co-occurrence is split into multiple continuous temporal feature segments based on the timestamp of the feature information. Each temporal feature segment corresponds to a specific stage in the defect evolution process.

[0099] The time-series processing module reads the timestamps from the cross-layer feature symbiosis, with timestamps accurate to the millisecond level. Based on the chronological order of the timestamps, the cross-layer feature symbiosis is divided into BBB time-series feature segments, with each segment having a time interval of CCC seconds. Each time-series feature segment contains feature component data and corresponding timestamp information for that time period, corresponding to a specific stage in the defect evolution process, such as the initial stage, development stage, or stable stage.

[0100] Step S142: Input each time series feature segment into the loop operation unit of the time series processing module in chronological order. Extract the feature change information in each time series feature segment through the activation operation of the hidden layer to obtain the time series change features. The feature change information includes the increase or decrease of feature components and the change in intensity.

[0101] The recurrent computation unit employs a Long Short-Term Memory (LSTM) network, with each temporal feature segment input into the network sequentially. The network's hidden layers extract feature change information through activation operations (using the ReLU function), specifically manifested as increases or decreases in the numerical values ​​and intensity changes of feature components. Each temporal feature segment corresponds to a temporal change feature with dimensions [1, DDD], where DDD represents the number of neurons in the hidden layer. The temporal change feature contains dynamic changes in the defective features within that stage.

[0102] Step S143: Establish the correlation between adjacent temporally changing features by comparing the overlap of feature components, forming a feature association chain. Perform path filtering operation on the feature association chain, and retain the association path that conforms to the evolutionary logic based on the consistency of feature change trends to obtain an effective evolutionary path.

[0103] The overlap ratio comparison operation of feature components calculates the intersection ratio of feature components between adjacent temporally changing features; a higher overlap ratio indicates a stronger correlation. Strongly correlated feature pairs are selected based on the overlap ratio threshold, forming feature association chains. The path filtering operation analyzes these feature association chains, retaining association paths with consistent feature change trends to obtain effective evolutionary paths.

[0104] Step S1431: Extract all associated path segments in the feature association chain through node parsing operations on the feature association chain. Each associated path segment consists of two adjacent temporal change features and their connection relationship.

[0105] The node parsing operation traverses all nodes in the feature association chain, with each node corresponding to a temporal change feature. For each node, the next node directly connected to it is extracted, forming an association path segment. Each association path segment contains the preceding temporal change feature, the following temporal change feature, and the association strength value connecting the two.

[0106] Step S1432: Calculate the gradient of change of features between two consecutive time-series changes in each associated path segment through feature change gradient calculation. The gradient of change reflects the rate and direction of increase or decrease of feature components.

[0107] For each associated path segment, calculate the difference between the subsequent temporal change features and the preceding temporal change features, and then divide this difference by the difference between the subsequent timestamp and the preceding timestamp to obtain the feature change gradient. The change gradient is in vector form, with each component corresponding to the rate and direction of change of one feature dimension; positive values ​​indicate an increase, and negative values ​​indicate a decrease.

[0108] Step S1433: Compare the gradient changes of adjacent related path segments and analyze whether the trends of changes of the related path segments before and after are consistent. A consistent trend is characterized by the same direction of the gradient changes and the difference in rate within a preset range.

[0109] The comparison operation calculates the cosine of the gradient angle between adjacent related path segments and the rate difference. If the cosine of the gradient angle is greater than the preset threshold EEE (indicating the same direction) and the rate difference is less than the preset threshold FFF (indicating stable rate change), then the trend of change is determined to be consistent.

[0110] Step S1434: Mark continuous associated path segments with consistent change trends to form candidate path sets. Each candidate path set consists of multiple continuous associated path segments with consistent trends.

[0111] All associated path segments are traversed, and continuous segments with consistent trends are marked as candidate path sets. Each candidate path set contains multiple associated path segments, arranged in chronological order to form a continuous evolutionary path.

[0112] Step S1435: Verify whether the attribute association of the feature components in each candidate path set conforms to the basic logic of defect evolution through feature attribute association operation. The logical verification is based on the source level and functional attributes of the feature components.

[0113] Feature attribute association operations verify whether the feature associations in the candidate path set conform to physical laws based on the source level of the feature components (such as surface circuit layer, internal via layer) and functional attributes (such as conductivity, insulation). For example, changes in internal via characteristics should precede changes in surface circuit characteristics, and changes in dielectric characteristics should affect the feature changes of both vias and circuits.

[0114] Step S1436: Eliminate the candidate path set that fails the logical verification, retain the candidate path set that passes the logical verification, form a preliminary valid path, perform a completeness analysis operation on the preliminary valid path, check whether the preliminary valid path covers all stages of defect evolution, so that the preliminary valid path contains the complete evolution process from the initial feature to the final feature, supplement the missing evolution stage related path segments in the preliminary valid path, generate the missing related path segments through interpolation operation of adjacent stage features, so that the preliminary valid path forms a continuous structure.

[0115] Candidate paths that fail logical verification are removed, and those that pass verification are retained to form preliminary valid paths. Integrity analysis checks whether the preliminary valid paths include the initial, development, stable, and manifest stages of defect evolution. If missing stages exist, missing related path segments are generated through interpolation of features from adjacent stages.

[0116] For example, step S1436-1: Perform stage traversal operation on the preliminary valid path, check each evolution stage in the preliminary valid path one by one, compare the timestamp marks of adjacent evolution stages, determine the stage interval where the timestamp interval exceeds the preset range, and use the stage interval as the associated path segment of the missing evolution stage.

[0117] The stage traversal operation checks each evolution stage in the initially valid path in chronological order and calculates the timestamp interval between adjacent stages. If the timestamp interval is greater than a preset threshold GGG, it is determined that there is a missing evolution stage in that interval, and the associated path segment needs to be supplemented.

[0118] Step S1436-2: Extract the two adjacent temporal change features before and after the missing associated path segment, and determine the feature component parameters and timestamp markers of the preceding temporal change feature and the feature component parameters and timestamp markers of the following temporal change feature.

[0119] Extract the last temporal change feature before the missing interval (preceding feature) and the first temporal change feature after the missing interval (following feature), and record their feature component parameters and timestamps.

[0120] Step S1436-3: Calculate the time interval based on the timestamps of the preceding and following time series change features, determine the number of interpolation points based on the time interval, and establish a positive correlation between the number of interpolation points and the time interval; perform interpolation calculations on each feature component of the preceding and following time series change features through linear interpolation operations to generate feature component parameters corresponding to each interpolation point, and the interpolation calculation is based on the changing trends of the feature component parameters of the preceding and following time series change features.

[0121] The time interval is obtained by subtracting the preceding timestamp from the subsequent timestamp. The number of interpolation points is set to the time interval divided by the preset interpolation time unit HHH, rounded up. Linear interpolation is performed separately for each feature component, calculating intermediate parameter values ​​based on the parameters of the preceding and following feature components and the position of the interpolation points, maintaining a consistent trend of change with the preceding and following features.

[0122] Step S1436-4: Arrange the generated interpolation point feature component parameters in the order of timestamps to form an interpolation feature sequence. The timestamp of the interpolation feature sequence is located between the timestamp of the preceding time series change feature and the timestamp of the following time series change feature.

[0123] The timestamps of the interpolation points are distributed at equal intervals between the preceding and following timestamps, and the interpolation feature sequence is arranged in the order of the timestamps to form a continuous feature change sequence.

[0124] Step S1436-5: Perform trend consistency verification operation on the interpolated feature sequence, compare the change trend of the feature components of the interpolated feature sequence with the change trends of the preceding time series change features and the following time series change features, so that the change trend of the interpolated feature sequence is consistent with the overall evolution trend.

[0125] The trend consistency verification operation calculates the gradient of the interpolated feature sequence and compares it with the gradients of the preceding and following features to ensure that the direction is consistent and the rate is stable.

[0126] Step S1436-6: If the trend consistency verification fails, adjust the parameters of the interpolation operation and recalculate the interpolation until the generated interpolated feature sequence meets the trend consistency requirements.

[0127] If the verification fails, adjust the number of interpolation points or use a nonlinear interpolation algorithm (such as cubic spline interpolation) to regenerate the interpolation feature sequence and perform verification again.

[0128] Step S1436-7: Convert the interpolated feature sequence that has passed the trend consistency verification into the associated path fragment format to form a supplementary associated path fragment for missing associated path fragments.

[0129] The interpolated feature sequence is converted into an associated path fragment format, which includes preceding features, subsequent features, and connection relationships, and is consistent with the original associated path fragment format.

[0130] Step S1436-8: Insert the supplementary associated path fragment into the corresponding missing position in the initial valid path. Through path connection operation, seamlessly connect the supplementary associated path fragment with the preceding associated path fragment and the following associated path fragment, so that the connected path forms a continuous structure. Perform a continuity verification operation on the supplemented complete path, traverse all stages and associated path fragments in the supplemented complete path, and confirm that the timestamps of all adjacent associated path fragments are continuous and the change trend is consistent, thus forming a valid evolution path.

[0131] The missing segments of the associated path are inserted to fill in the gaps in the initially valid path. Path connection operations ensure that timestamps are continuous and their change trends are consistent. Continuity verification operations traverse all stages and segments to confirm that the path is complete and continuous.

[0132] Step S1437: Perform a secondary trend verification operation on the supplemented path to confirm the consistency of the change trend between the supplemented related path segments and the original related path segments, so that the supplemented related path segments are consistent with the overall evolution logic. Integrate all related path segments that have passed the secondary verification to form an effective evolution path. The effective evolution path fully reflects the continuous evolution process of the defect from the initial state to the manifest state.

[0133] The secondary trend verification operation performs an overall trend analysis on the supplemented complete path to ensure that the change trend of the supplemented segment is consistent with that of the original segment and conforms to the overall logic of defect evolution. The verified related path segments are then integrated to form the final effective evolution path.

[0134] Step S144: Perform an identifier assignment operation on each temporal change feature in the effective evolution path. Based on the temporal position of the temporal change feature and the parameter features of the temporal change feature, assign a unique evolutionary stage identifier to each temporal change feature to obtain the stage label feature.

[0135] The identifier allocation operation assigns a unique evolutionary stage identifier to each temporal variation feature based on its timestamp position and parameter characteristics (such as the maximum value and rate of change of feature components). The identifier is in string format, containing a timestamp abbreviation and a parameter feature hash value to ensure uniqueness. The stage marker feature contains the temporal variation feature data and the corresponding evolutionary stage identifier.

[0136] Step S145: The trajectory reconstruction submodule of the time-series processing module converts the marker features of each stage into spatial trajectory points on a two-dimensional plane through coordinate mapping operations. The coordinate position of each spatial trajectory point corresponds to the feature state of an evolution stage. Furthermore, the spatial coordinate points are connected sequentially in time through line connection operations to form a preliminary defect evolution trajectory line. The direction of the preliminary defect evolution trajectory line is consistent with the direction of defect evolution.

[0137] The coordinate mapping operation uses principal component analysis to reduce the dimensionality of the high-dimensional stage marker features to two-dimensional space, obtaining the horizontal and vertical coordinates of the spatial trajectory points. The line connection operation connects the trajectory points in chronological order using directed line segments. The direction of the line segment represents the defect evolution direction, and the length of the line segment represents the evolution speed.

[0138] Step S146: Perform smoothing operation on the preliminary defect evolution trajectory to form a continuous defect evolution trajectory. Mark the feature association information of each stage on the continuous defect evolution trajectory. The feature association information includes the interaction mode of feature components at different levels and the change parameters of association strength. The marking information corresponds one-to-one with the spatial coordinates of the continuous defect evolution trajectory.

[0139] The smoothing process uses a moving average filtering algorithm to smooth the initial trajectory line and eliminate noise interference. Annotation information is added next to the corresponding coordinate points on the trajectory line in the form of text boxes, including feature component IDs, interaction methods (such as enhancement and suppression), and correlation strength change parameters.

[0140] Step S147: Transmit the labeled continuous defect evolution trajectory lines to the network topology generation module for format normalization calculation, convert the continuous defect evolution trajectory lines and labeling information into a standard network topology format, and form a defect trajectory reconstruction network topology.

[0141] The format normalization operation converts the trajectory lines and annotation information into network topology data conforming to the GraphML standard, containing node (corresponding to evolution stage) and edge (corresponding to evolution path) elements. Node attributes include coordinates, timestamps, and feature states; edge attributes include direction, length, and association strength. The resulting defect trajectory reconstructed network topology is stored in file format and can be read and visualized by network analysis tools.

[0142] Step S150: The network topology is reconstructed by parsing the defect trajectory through the network topology generation module of the deep learning defect detection model, the initial generation level and diffusion path information of the defect are located, and a detection report containing cross-layer feature co-occurrence association and defect evolution information is generated.

[0143] The information extraction submodule of the network topology generation module traverses the nodes of the network topology to reconstruct the network topology based on the defect trajectory, identifying the layer corresponding to the node with the earliest timestamp as the initial defect generation layer. It calculates the length, direction, and velocity parameters of the diffusion path based on the spatial coordinate sequence and timestamp sequence of the nodes. It integrates cross-layer feature co-occurrence association information and defect evolution information to generate a structured inspection report, including the defect ID, initial layer, diffusion path, feature association list, and evolution stage division. The inspection report is output in both JSON and PDF formats, stored in a specified directory, and pushed to the production management system via an API interface.

[0144] Figure 2 The diagram illustrates the hardware structure of an HDI communication board defect detection system 100 incorporating deep learning, as provided in an embodiment of the present invention. Figure 2 As shown, the HDI communication board defect detection system 100, which incorporates deep learning, may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0145] Machine-readable storage medium 120 can store data and / or instructions. In some embodiments, machine-readable storage medium 120 can store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 can store data and / or instructions used by the deep learning-integrated HDI communication board defect detection system 100 to execute or use in order to complete the exemplary methods described in this invention. In a specific implementation, one or more processors 110 execute the computer-executable instructions stored in machine-readable storage medium 120, causing processor 110 to execute the deep learning-integrated HDI communication board defect detection method as described in the above method embodiments. Processor 110, machine-readable storage medium 120, and communication unit 140 are connected via bus 130, and processor 110 can be used to control the transmission and reception operations of communication unit 140. The specific implementation process of processor 110 can be found in the various method embodiments executed by the deep learning-integrated HDI communication board defect detection system 100 described above, and their implementation principles and technical effects are similar, so they will not be repeated here.

[0146] Furthermore, embodiments of the present invention also provide a readable storage medium containing computer-executable instructions. When a processor executes the computer-executable instructions, the above-described HDI communication board defect detection method incorporating deep learning is implemented.

[0147] It should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof.

Claims

1. A method for defect detection of HDI communication board combined with deep learning, characterized in that, The method includes: Synchronously acquire multi-layer structure images of the HDI communication board, including surface circuit images, internal via images, and interlayer dielectric images; A deep learning defect detection model is constructed, which integrates a feature extraction module, a co-occurrence association module, a temporal processing module, and a network topology generation module. The multi-layer structured image is input into the feature extraction module of the deep learning defect detection model. The defect-related features of each layer of the image are extracted through convolution operation within the feature extraction module. The correlation between defect-related features is calculated by the co-occurrence correlation module to generate a cross-layer feature co-occurrence containing cross-layer feature correlation information. The time-series processing module of the deep learning defect detection model processes cross-layer feature co-occurrences, reconstructs the defect evolution path, and forms a defect trajectory reconstruction network topology. The defect trajectory reconstruction network topology presents the evolution process of the defect from the initial state to the manifest state. The network topology generation module of the deep learning defect detection model analyzes the defect trajectory to reconstruct the network topology, locates the initial generation level and diffusion path information of the defect, and generates a detection report containing cross-layer feature co-occurrence association and defect evolution information. 2.The HDI communication board defect detection method with deep learning of claim 1, wherein, The construction of the deep learning defect detection model includes: The overall architecture for constructing a deep learning defect detection model adopts a modular serial structure, which sequentially includes an input interface layer, a feature extraction module, a co-occurrence association module, a temporal processing module, a network topology generation module, and an output interface layer. Within the feature extraction module, a multi-size convolutional kernel group is set up for extracting features from multi-layer structure images. The multi-size convolutional kernel group contains convolutional kernels of specific sizes corresponding to surface line images, internal via images, and interlayer medium images, respectively. The size of each specific-size convolutional kernel is set based on the pixel distribution statistics of the features in the corresponding layer image. A feature preprocessing submodule is set in the feature extraction module. The feature preprocessing submodule performs pixel normalization operation on the input multi-layer structure image and linearly transforms the pixel values ​​of each image to a preset numerical range. A feature cross-operation unit is set in the symbiotic association module to perform channel-by-channel feature comparison operation on the input defect-related features of different levels. The channel-by-channel feature comparison operation calculates the association relationship between features of different levels by analyzing the synchronous change trend of feature response values. A weight allocation submodule is set in the symbiotic association module. The weight allocation submodule assigns weights to each feature component based on the correlation strength value between feature components. A loop operation unit is set in the time series processing module to receive feature sequences arranged in order of timestamps, and to establish the correlation between features before and after in the feature sequence through the internal state transmission of the loop operation unit; A trajectory reconstruction submodule is set in the time series processing module. The trajectory reconstruction submodule performs interpolation operations on the serialized feature data and generates intermediate feature states based on the changing gradient between adjacent features in order to reconstruct the continuous defect evolution path. A network topology drawing unit is set up in the network topology generation module to perform coordinate mapping calculations, map each evolution stage in the defect evolution path to spatial coordinate points, and connect the spatial coordinate points in chronological order to generate network topology data. An information extraction submodule is set in the network topology generation module. The information extraction submodule traverses the nodes in the network topology data, identifies the level corresponding to the node with the earliest timestamp as the initial generation level of the defect, and calculates the diffusion path parameters based on the spatial coordinate sequence and timestamp sequence of the node. The deep learning defect detection model connects the input interface layer, feature extraction module, co-occurrence association module, temporal processing module, network topology generation module, and output interface layer sequentially through the internal data bus of the deep learning defect detection model, thus forming the deep learning defect detection model. 3.The HDI communication board defect detection method with deep learning of claim 1, wherein, The feature extraction module of the deep learning defect detection model inputs multi-layer structured images into the model, extracts defect-related features of each layer through convolution operations within the feature extraction module, and calculates the correlation between defect-related features through a co-occurrence correlation module to generate a cross-layer feature co-occurrence containing cross-layer feature correlation information, including: The multi-layer structure image is imported into the deep learning defect detection model through the input interface layer according to the categories of surface line image, internal via image, and interlayer medium image. The image data is converted into a feature input format that can be recognized by the deep learning defect detection model through format conversion operation. The converted surface line image is input into the surface line feature extraction subunit of the feature extraction module. Through continuous sliding operation of multi-size convolution kernels, the edge contour features and line distribution features in the surface line image are extracted region by region. The surface line features are obtained after multiple rounds of convolution superposition operation. The converted internal through-hole image is input into the internal through-hole feature extraction subunit of the feature extraction module. Through multi-directional scanning convolution operation is used to cover various angles of the through-hole region to extract the aperture morphology features and through-hole connectivity features in the internal through-hole image. After dimensional integration operation, the internal through-hole features are obtained. The converted interlayer medium image is input into the interlayer medium feature extraction subunit of the feature extraction module. Through the adaptation operation of the stride adjustable convolution kernel, the bonding density feature and medium distribution feature in the interlayer medium image are extracted. The interlayer medium feature is obtained through feature aggregation operation. The surface circuit features, internal via features, and interlayer dielectric features are synchronously transmitted to the feature cross-operation unit of the symbiotic association module through the internal data bus of the deep learning defect detection model. The surface circuit features and internal via features are fused channel by channel. The feature response values ​​of each channel are compared and the feature components with synchronously changing response values ​​in the two types of features are retained to obtain the circuit via association features. The circuit via association features and interlayer dielectric features are fused across channels to integrate the feature components with synchronous changing trends to obtain the cross-layer association features. The weight allocation submodule of the symbiotic association module performs weight adjustment calculations on the cross-layer association features. The weight adjustment calculations assign corresponding weight parameters to each feature component based on the association strength values ​​between features, thereby obtaining weighted cross-layer association features. The feature integration and operation unit of the symbiotic association module performs dimensional unification operation on the weighted cross-layer association features, performs association solidification operation on the unified dimension features, and arranges the scattered symbiotic relationships in an orderly manner according to the weight ratio of the feature components, forming a cross-layer feature symbiosis that includes symbiotic association information of surface circuit features, internal via features and interlayer medium features. 4.The HDI board defect detection method with deep learning of claim 1, wherein, The temporal processing module of the deep learning defect detection model processes cross-layer feature co-occurrences, reconstructs the defect evolution path, and forms a defect trajectory reconstruction network topology, including: The temporal processing module of the deep learning defect detection model splits the cross-layer feature co-occurrence into multiple continuous temporal feature segments based on the timestamp of the feature information. Each temporal feature segment corresponds to a specific stage in the defect evolution process. Each time-series feature segment is input into the loop operation unit of the time-series processing module in chronological order. The feature change information in each time-series feature segment is extracted through the activation operation of the hidden layer to obtain the time-series change features. The feature change information includes the increase or decrease of feature components and the change in intensity. By comparing the overlap of feature components, the correlation between adjacent temporally changing features is established, forming a feature association chain. Path filtering operation is performed on the feature association chain, and the association path that conforms to the evolutionary logic is retained based on the consistency of feature change trends, thus obtaining an effective evolutionary path. For each temporal change feature in the effective evolutionary path, an identifier assignment operation is performed. Based on the temporal position of the temporal change feature and the parameter characteristics of the temporal change feature, a unique evolutionary stage identifier is assigned to each temporal change feature, thus obtaining the stage label feature. The trajectory reconstruction submodule of the time-series processing module converts the marked features of each stage into spatial trajectory points on a two-dimensional plane through coordinate mapping operations. The coordinate position of each spatial trajectory point corresponds to the feature state of an evolution stage. Furthermore, the spatial trajectory points are connected sequentially in time through line connection operations to form a preliminary defect evolution trajectory line. The direction of the preliminary defect evolution trajectory line is consistent with the direction of defect evolution. The initial defect evolution trajectory is smoothed to form a continuous defect evolution trajectory. Feature association information of each stage is marked on the continuous defect evolution trajectory. The feature association information includes the interaction mode of feature components at different levels and the variation parameters of association strength. The marked information corresponds one-to-one with the spatial coordinates of the continuous defect evolution trajectory. The labeled continuous defect evolution trajectory lines are transmitted to the network topology generation module for format normalization calculation, which converts the continuous defect evolution trajectory lines and labeling information into a standard network topology format, forming a defect trajectory reconstructed network topology. 5.The HDI board defect detection method with deep learning of claim 2, wherein, The feature extraction module includes a multi-size convolutional kernel group for extracting features from multi-layer structure images. This multi-size convolutional kernel group contains convolutional kernels of specific sizes corresponding to surface circuit images, internal via images, and interlayer dielectric images, respectively. The size of each specific-size convolutional kernel is set based on the pixel distribution statistics of the features in the corresponding layer image, including: Analyze the pixel distribution of line features in the surface line image, determine the average pixel spacing of the surface line features in the image, and preliminarily set the convolution kernel size range corresponding to the surface line feature extraction based on the average pixel spacing; By training surface line image samples in the deep learning defect detection model training sample set, the initially set convolution kernel size is fine-tuned. Convolution kernels of different sizes are applied to feature extraction of surface line image samples. The completeness of the extracted line features is compared, and the convolution kernel size that can completely cover the line features is selected. Analyze the pixel distribution of the via features in the internal via image, determine the average pixel diameter of the via features in the image, and preliminarily set the convolution kernel size range corresponding to the internal via feature extraction based on the average pixel diameter; By training the deep learning defect detection model on internal through-hole image samples in the sample set, the initially set convolution kernel size is fine-tuned. Convolution kernels of different sizes are applied to feature extraction of internal through-hole image samples. The completeness of the extracted through-hole features is compared, and the convolution kernel size that can completely cover the through-hole features is selected. Analyze the pixel distribution of medium features in the interlayer medium image, determine the average pixel thickness of the medium features in the image, and preliminarily set the convolution kernel size range corresponding to the extraction of interlayer medium features based on the average pixel thickness; By training the deep learning defect detection model on interlayer medium image samples in the sample set, the initially set convolution kernel size is fine-tuned. Convolution kernels of different sizes are applied to feature extraction of interlayer medium image samples. The completeness of the extracted medium features is compared, and the convolution kernel size that can completely cover the medium features is selected. The selected surface line feature extraction convolution kernels, internal via feature extraction convolution kernels, and interlayer medium feature extraction convolution kernels are integrated into a multi-size convolution kernel group for the feature extraction module. Each group of convolution kernels corresponds to a unique hierarchical image feature extraction task. The stride parameters of the convolution kernels are set as follows: the stride parameter of the surface line feature extraction convolution kernel is set to a preset ratio of the average pixel spacing of the line feature; the stride parameter of the internal via feature extraction convolution kernel is set to a preset ratio of the average pixel diameter of the via feature; and the stride parameter of the interlayer medium feature extraction convolution kernel is set to a preset ratio of the average pixel thickness of the medium feature. Set the number of convolution kernels parameter, determine the number of convolution kernels corresponding to the number of feature dimensions of each layer of the image, so that each feature dimension corresponds to a unique convolution kernel for extraction operation, and achieve one-to-one correspondence through parameter binding; The determined convolution kernel size, stride, and number parameters are entered into the parameter configuration file of the feature extraction module. The configuration parameters are then applied to the feature extraction process through parameter loading operations within the feature extraction module, thereby achieving the corresponding adaptation of the convolution kernel to the feature distribution density of each layer of the image. 6.The HDI communication board defect detection method with deep learning of claim 3, wherein, The weight allocation submodule of the symbiotic association module performs weight adjustment calculations on the cross-layer association features. Based on the association strength analysis results between features, it assigns corresponding weight parameters to each feature component to obtain weighted cross-layer association features, including: The weight allocation submodule of the symbiotic association module extracts all feature components in the cross-layer association features through feature parsing operations, and determines the source level and attribute information of each feature component; The correlation strength between any two feature components can be determined by calculating the Pearson correlation coefficient or covariance of the response values ​​between feature components. The correlation strength of each pair of feature components is quantified to obtain the feature correlation strength value. The magnitude of the feature correlation strength value directly reflects the tightness of the symbiotic relationship between feature components. Set weight allocation rules, divide multiple weight levels according to the distribution range of feature association strength values, and each weight level corresponds to a fixed weight parameter. The higher the association strength value, the larger the corresponding weight parameter. Based on the feature association strength value, a corresponding weight level is matched for each feature component, and then the initial weight parameter of each feature component is determined. The initial weight parameter is normalized and the weight parameters of all feature components are adjusted to the 0-1 range so that the sum of the weight parameters of each feature component is 1. In the weight allocation submodule, a weight verification unit is constructed. The weight verification unit performs the following operations: inputting the features of multi-layer structured image samples with initial weight parameters into the co-occurrence association module, calculating the weighted cross-layer association features of the samples and the association relationship calculation results based on them; comparing the association relationship calculation results with the real defect association relationships of the pre-labeled samples, and calculating the matching degree between the two. The initial weight parameters are adjusted based on the verification results. The weight parameters of the feature components whose matching degree does not meet the standard are fine-tuned. The adjusted weight parameters are then bound to the corresponding feature components to form a set of feature components with weight labels. Through the weight allocation submodule of the symbiotic association module, the set of feature components with weight labels is integrated into a feature vector to obtain a weighted cross-layer association feature. The weight ratio of each component in the weighted cross-layer association feature corresponds to the tightness of the symbiotic relationship between the features. 7.The HDI board defect detection method with deep learning of claim 4, wherein, The path filtering operation on the feature association chain, based on the consistency of feature change trends, retains association paths that conform to evolutionary logic to obtain effective evolutionary paths, including: All associated path segments in the feature association chain are extracted through node parsing operations on the feature association chain. Each associated path segment consists of two adjacent temporal variation features and their connection relationship. The gradient of change of features between two consecutive time-series changes in each associated path segment is calculated by the gradient of feature change. The gradient of change reflects the rate and direction of increase or decrease of feature components. Compare the gradient changes of adjacent related path segments and analyze whether the trends of changes of related path segments before and after are consistent. A consistent trend is characterized by the same direction of the gradient changes and the rate difference being within a preset range. Mark consecutive related path segments with consistent trends of change to form candidate path sets. Each candidate path set consists of multiple consecutive related path segments with consistent trends. The attribute association operation is used to verify whether the attribute association of the feature components in each candidate path set conforms to the basic logic of defect evolution. The logical verification is based on the source hierarchy and functional attributes of the feature components. The candidate path set that fails the logical verification is removed, and the candidate path set that passes the logical verification is retained to form a preliminary effective path. The preliminary effective path is subjected to integrity analysis operation to check whether the preliminary effective path covers all stages of defect evolution, so that the preliminary effective path contains the complete evolution process from the initial feature to the final feature. The missing evolution stage related path segments in the preliminary effective path are supplemented, and the missing related path segments are generated by interpolation operation of adjacent stage features, so that the preliminary effective path forms a continuous structure. A secondary trend verification operation is performed on the supplemented path to confirm the consistency of the change trend between the supplemented related path segments and the original related path segments. This ensures that the supplemented related path segments are consistent with the overall evolution logic. All related path segments that have passed the secondary verification are integrated to form an effective evolution path. The effective evolution path fully reflects the continuous evolution process of the defect from the initial state to the manifest state. 8.The HDI board defect detection method with deep learning of claim 5, wherein, The process involves fine-tuning the initially set convolutional kernel size using surface line image samples trained on the deep learning defect detection model training sample set. Different sized convolutional kernels are applied to feature extraction from the surface line image samples. The completeness of the extracted line features is compared, and the convolutional kernel size that can fully cover the line features is selected, including: Surface line image samples are selected from the training sample set of the deep learning defect detection model. The surface line image samples contain surface line maps with different line distribution densities and different line widths. The pixel distribution of line features in the surface line images is analyzed to determine the average pixel spacing of surface line features in the image. Based on the average pixel spacing, the convolution kernel size range corresponding to the surface line feature extraction is initially set. The initially set convolution kernel size range is divided into multiple continuous size gradients, and each size gradient corresponds to a specific convolution kernel size parameter. For each size gradient, a temporary feature extraction unit is constructed. Other parameters of the temporary feature extraction unit are kept fixed, and only the kernel size parameter is changed. The selected surface line image samples are input into each temporary feature extraction unit one by one. The surface line features in the surface line image samples are extracted by the convolution operation in the temporary feature extraction unit, and the line feature extraction results of each surface line image sample under different convolution kernel sizes are obtained. The integrity calculation of the line contour is used to evaluate whether the line contour in each line feature extraction result is complete, without any breaks or missing parts. The line distribution coverage calculation is used to evaluate whether the line distribution in each line feature extraction result completely corresponds to the actual line distribution in the surface line image sample, with no missing lines and no false lines. By combining the results of the line contour integrity calculation and the results of the line distribution coverage calculation, the line feature extraction results under each convolution kernel size are quantitatively scored. The higher the score, the better the line feature extraction integrity. The average score of all surface line image samples at each convolutional kernel size was statistically analyzed. The average scores of different convolutional kernel sizes were compared, and the convolutional kernel size range with the highest average score was selected. Within the selected kernel size range, the size gradient is further subdivided, and the above steps of surface line image sample extraction, line feature integrity evaluation and scoring statistics are repeated to obtain the optimal kernel size. The optimal convolution kernel size is determined as the final convolution kernel size corresponding to surface line feature extraction, and is entered into the parameter configuration file of the feature extraction module as a fixed operation parameter for surface line feature extraction of the feature extraction module. 9.The HDI communication board defect detection method with deep learning of claim 6, wherein, The weight verification unit is constructed in the weight allocation submodule. The weight verification unit performs the following operations: inputting the features of multi-layer structured image samples with initial weight parameters into the co-occurrence association module, and calculating the weighted cross-layer association features of the samples and the association relationship calculation results based on them. The calculated correlation result is compared with the actual defect correlation in the pre-labeled samples to calculate the degree of matching, including: Multi-layer structure image samples are selected from the training sample set of the deep learning defect detection model. The multi-layer structure image samples contain annotation information. The annotation information records whether there is a predefined correlation between surface circuit features, internal via features and interlayer medium features, as well as the type identifier of the correlation. The selected multilayer structure image samples are input into the feature extraction module of the deep learning defect detection model according to the categories of surface circuit images, internal via images, and interlayer medium images. The surface circuit features, internal via features, and interlayer medium features of the multilayer structure image samples are extracted through convolution operations in the feature extraction module. The extracted multi-layer image sample features are transmitted to the co-occurrence association module, and the initial weight parameters are applied to perform weight adjustment operations on the multi-layer image sample features to obtain the weighted cross-layer association features of the multi-layer image samples. The feature cross-operation unit of the symbiotic association module performs symbiotic relationship mining operations on the weighted cross-layer association features of multi-layer structured image samples to obtain the mining association relationship of multi-layer structured image sample features; The correlation matching operation compares the feature mining correlation of multi-layer structure image samples with the actual defect correlation of multi-layer structure image samples one by one, and counts the number of successfully matched correlations. The ratio of the number of successfully matched associations to the total number of actual defect associations is calculated to obtain the matching degree value. The calculated matching degree value is compared with the matching degree threshold. If the matching degree value is higher than or equal to the matching degree threshold, the initial weight parameter is deemed to have passed the verification. If the matching degree value is lower than the matching degree threshold, the unmatched associations are marked. The feature components corresponding to the unmatched associations are analyzed to determine the weight parameters that caused the matching failure. The verification results and the analysis report of unmatched relationships are fed back to the weight allocation submodule of the symbiotic relationship module.

10. A deep learning combined HDI communication board defect detection system, characterized in that, The HDI communication board defect detection system incorporating deep learning includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions, or code, and the processor is used to run the programs, instructions, or code in the memory to implement the HDI communication board defect detection method incorporating deep learning as described in any one of claims 1-9.

Images