A heat dissipation substrate defect detection method and system based on deep learning

Abstract

The embodiment of the present application provides a kind of based on deep learning's heat dissipation substrate defect detection method and system, it is related to the technical field of heat dissipation substrate defect detection technology.Its method includes: obtaining the three-dimensional point cloud data of the heat dissipation substrate to be detected;The three-dimensional point cloud data is preprocessed to generate standardized input point cloud;The geometric features of the input point cloud are extracted and fused by the pre-trained defect detection model at multiple scales to obtain defect class prediction results.Through the present application, the problem of low complex defect recognition accuracy is solved, and the effect of improving the complex defect recognition accuracy is achieved.

Description

Technical Field

[0001] The present invention relates to the field of heat dissipation substrate defect detection technology, and more specifically, to a heat dissipation substrate defect detection method and system based on deep learning. Background Technology

[0002] In modern electronic and power equipment manufacturing, heat dissipation substrates are key components that ensure the temperature stability of core components, and their manufacturing precision directly affects the reliability and service life of the entire system. However, during the production process, due to the combined effects of material properties, processing technology, and environmental factors, various geometric defects such as bending, corrosion, and scratches inevitably occur on the surface of the heat dissipation substrate.

[0003] Traditional defect detection primarily relies on manual visual inspection or 2D machine vision technology. The former is inefficient and highly subjective, while the latter, due to the inherent lack of depth information in 2D images, struggles to accurately quantify and identify geometric distortions in three-dimensional space, such as flatness deviations and base warping, and is highly susceptible to interference from variations in industrial lighting and surface textures. 3D point cloud technology, by directly acquiring the spatial coordinates of an object's surface, provides an ideal data carrier for solving these problems. However, the inherent disorder and unstructured nature of point cloud data poses challenges to feature extraction for deep learning models.

[0004] Existing 3D point cloud processing networks struggle to effectively capture both microscopic local geometric details (such as minute scratches) and perceive macroscopic global contextual information (such as overall curvature) within a single network architecture. This limits their ability to identify complex defects. In particular, since the heat dissipation substrate exhibits both large-scale overall warping and extremely small local scratches, existing technologies struggle to effectively decouple and accurately capture these two types of features within the same model. Summary of the Invention

[0005] This invention provides a method and system for detecting defects in a heat dissipation substrate based on deep learning, which at least solves the problem of low accuracy in identifying complex defects in related technologies.

[0006] According to an embodiment of the present invention, a method for detecting defects in a heat dissipation substrate based on deep learning is provided, comprising: Acquire the three-dimensional point cloud data of the heat dissipation substrate to be tested; The three-dimensional point cloud data is preprocessed to generate a standardized input point cloud; The geometric features of the input point cloud are extracted and fused at multiple scales using a pre-trained defect detection model to obtain defect category prediction results.

[0007] In an exemplary embodiment, the defect detection model performs in-depth correction and refinement of the features of the input point cloud through a dual offset attention module; wherein, the dual offset attention module includes a global offset generation branch and a local offset generation branch, the global offset generation branch is used to calculate the global feature offset representing the global context bias, and the local offset generation branch is used to calculate the local feature offset representing the local geometric distortion.

[0008] In an exemplary embodiment, the global offset generation branch generates global context features based on self-attention calculation of the input point cloud features, and determines the global feature offset based on the difference between the input point cloud features and the global context features; The local offset generation branch processes the input point cloud features through depthwise separable convolution to give them a local receptive field, calculates local geometric features based on the processed features, and then determines the local feature offset based on the difference between the input point cloud features and the local geometric features.

[0009] In one exemplary embodiment, before inputting the standardized input point cloud into the defect detection model, the method further includes: The standardized input point cloud is transformed into geometric graph structure data; wherein, the data points in the input point cloud constitute the vertices of the geometric graph structure, and the edges between the vertices are determined based on the spatial proximity relationship between the data points.

[0010] In one exemplary embodiment, the defect detection model is a graph neural network model; The graph neural network model processes the geometric graph structure data through a message passing mechanism. Each vertex aggregates the feature information of its neighboring vertices to iteratively update its own feature representation, thereby enabling the extraction of local and global defect features.

[0011] In one exemplary embodiment, the method further includes: Calculate initial node features for each vertex of the geometric graph structure, the initial node features including the three-dimensional coordinates of the data point and the normal vector or curvature calculated based on its local neighborhood; Calculate initial edge features for each edge of the geometric graph structure, the initial edge features including the relative position vector or Euclidean distance between the two vertices connected by the edge.

[0012] In one exemplary embodiment, preprocessing the three-dimensional point cloud data includes: A radius filtering algorithm is used to remove background noise points and outliers from the 3D point cloud data; The farthest point sampling algorithm is used to uniformly sample the three-dimensional point cloud data to a preset number of points; The sampled 3D point cloud data is normalized so that its coordinates lie within a unit sphere.

[0013] According to another embodiment of the present invention, a deep learning-based defect detection system for heat dissipation substrates is provided, comprising: The data acquisition module is used to acquire the three-dimensional point cloud data of the heat dissipation substrate to be tested; The data preprocessing module is used to preprocess the three-dimensional point cloud data to generate a standardized input point cloud. The defect detection module is used to extract and fuse the geometric features of the input point cloud at multiple scales using a pre-trained defect detection model to obtain defect category prediction results.

[0014] According to yet another embodiment of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, wherein the computer program is configured to perform the steps in any of the above method embodiments when executed.

[0015] According to yet another embodiment of the present invention, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0016] This invention captures global contextual deformation and local geometric details in parallel using dual offset attention modules, and performs feature correction through residual learning, effectively identifying macroscopic and microscopic defects. Simultaneously, by transforming point clouds into geometric graph structures and utilizing the message passing mechanism of graph neural networks, features are propagated and aggregated in the graph domain, thereby understanding the topological relationships of point clouds at a higher dimension. This results in a strong ability to identify complex defects defined by inter-point relationships, thus solving the problem of low accuracy in identifying complex defects and improving the accuracy of complex defect identification. Attached Figure Description

[0017] Figure 1 This is a flowchart of a deep learning-based defect detection method for a heat dissipation substrate according to an embodiment of the present invention; Figure 2 This is a structural block diagram of a neural network model according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0019] In the following description, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0020] Furthermore, in this application, directional terms such as "upper," "lower," "left," and "right" may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and may change accordingly depending on the orientation of the components in the accompanying drawings.

[0021] In this application, unless otherwise expressly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a direct connection or an indirect connection through an intermediate medium. Furthermore, the term "coupled" can refer to an electrical connection that enables signal transmission.

[0022] As used herein, “about,” “approximately,” or “approximately” includes the stated value and the average value within an acceptable range of deviation from the given value, wherein the acceptable range of deviation is determined by a person skilled in the art taking into account the measurement under discussion and the error associated with the measurement of the given quantity (i.e., the limitations of the measurement system).

[0023] Example 1 This embodiment provides a deep learning-based method for detecting defects in heat dissipation substrates. By constructing a parallel dual-branch attention module, this module decouples and specializes the long-distance global context information and microscopic local geometric details, respectively. After fusing the two types of information, the original features are corrected and refined using a residual learning paradigm. This solves the technical problem that existing 3D point cloud detection networks cannot simultaneously take into account macroscopic deformation and microscopic defects, and achieves the beneficial effect of improving the accuracy of complex defect detection.

[0024] like Figure 1 The diagram illustrates the overall flow of a deep learning-based defect detection method for heat dissipation substrates provided in an embodiment of this application. The method may specifically include the following steps: S100: Acquire the three-dimensional point cloud data of the heat dissipation substrate to be tested.

[0025] This step is performed using a 3D data acquisition system that integrates non-contact 3D scanning hardware to acquire data on the surface morphology of heat dissipation substrates on industrial production lines.

[0026] For example, a 3D data acquisition system may include a line laser scanner that moves along a predetermined straight path parallel to the long axis of the substrate, while a rotary stage drives the heat dissipation substrate to rotate about its central vertical axis at a constant angular velocity. The scanner's laser emitter emits lasers at a frequency of 1000 laser lines per second, and the contour formed by each laser line on the substrate surface is captured by a camera and resolved into 2000 data points; thus, after a complete 360-degree scan taking 5 seconds, a raw point cloud dataset containing 10,000,000 data points is generated. Each data point... Structured into a data tuple The tuple contains its three-dimensional coordinates in the world coordinate system (in millimeters) and an 8-bit unsigned integer value representing the intensity of laser reflection. .

[0027] S200: Preprocesses 3D point cloud data to generate standardized input point clouds.

[0028] Because the raw point cloud data acquired by S100 contains background noise, outliers, and inconsistent point counts and densities between samples, the preprocessing module performs a series of standardization operations to generate clean and well-organized point cloud data suitable for the fixed input dimensions of subsequent deep learning models. Specifically: S210: Employs a radius filtering algorithm to remove background noise points and outliers from 3D point cloud data.

[0029] For each data point in the point cloud, the preprocessing module counts the number of its neighboring points within a spherical neighborhood of a preset radius R (e.g., 5.0 mm). If the number is less than a preset threshold K (e.g., 10), the data point is identified as an outlier and removed from the dataset. This effectively removes sparse noise points caused by measurement errors or environmental interference.

[0030] S220: Employs the farthest point sampling algorithm to uniformly sample 3D point cloud data to a preset number of points.

[0031] In response to the denoised point cloud generated in the previous step, the preprocessing module enables the farthest point sampling (FPS) process, which iteratively selects the point farthest from the original point set that is currently selected from the sampled point set until the number of sampled points reaches a preset target value N (e.g., 2048). This sampling strategy ensures that the generated subset can maintain the geometry and spatial coverage of the original point cloud to the greatest extent.

[0032] S230: Normalize the sampled 3D point cloud data so that its coordinates lie within a unit sphere.

[0033] To eliminate differences in spatial location and absolute scale among different samples, the preprocessing module performs centering and scaling operations. First, centering is achieved by calculating the centroid of the point cloud and translating it to the origin (0,0,0). Then, based on the Euclidean distance (e.g., 85.0 mm) of the point cloud furthest from the origin after centering, the coordinates of all points are scaled proportionally. Thus, all point cloud samples input to the model are constrained within a unit sphere centered at the origin with a radius of 1, allowing the model to focus on learning relative geometric features related to defects.

[0034] S300: Construct the training dataset and the test dataset.

[0035] The data management module performs manual annotation, data augmentation, and dataset partitioning on the preprocessed point cloud. For example, quality control engineers assign labels to each point cloud sample based on predefined defect categories (e.g., {0: normal, 1: bent, 2: corroded, 3: scratch, 4: warped, 5: dent}). To address the class imbalance caused by the large number of normal samples compared to defective samples, this module expands the data volume for minority classes such as scratches by applying a series of geometric and noise transformations, including random rotation, translation, scaling, and adding Gaussian noise. Finally, the entire augmented dataset is randomly divided into training, validation, and test sets in an 8:1:1 ratio, and so on.

[0036] S400: Input the preprocessed training dataset into the model for training.

[0037] Through iterative optimization, the defect detection model learns a complex mapping relationship from the input standardized point cloud to its corresponding defect category.

[0038] Please refer to the following: Figure 2 and Figure 1 In this embodiment, the defect detection model is a hierarchical network, and its core component is a dual offset attention module, specifically: S420: Feature depth correction and refinement are performed through a dual offset attention module.

[0039] In a forward propagation path, an input feature tensor X of dimensions [B, N, C] is fed into the dual offset attention module. Within this module, two parallel processing branches are activated: Global offset generation branch: This branch is used to capture the macroscopic deformation of the point cloud. Internally, a standard self-attention mechanism is invoked to compute global contextual features. Subsequently, the LBR (Linear-BatchNorm-ReLU) network block is based on the input feature X and the global context feature. The difference between them generates the final global feature offset. .

[0040] Local offset generation branch: This branch is used to capture local, microscopic geometric details; its input features are first processed through a depthwise separable convolutional (DWConv) layer, which contains a 3x3 depthwise convolutional kernel and a pointwise convolutional kernel, thus assigning local receptive fields to the features and generating... and Based on this, through The Hadamard product between key features K ( , Generate a local attention map for the Hadamard product operation.

[0041] The attention map was then fed into a nonlinear gating network. The network consists of a first fully connected layer (FC1) whose dimension is compressed from C to C / 2, a Swish activation function, a second fully connected layer (FC2) whose dimension is restored from C / 2 to C, and a Tanh activation function cascaded sequentially.

[0042] The output of the gating network is filtered by a learnable temperature coefficient. The weights are adjusted and normalized using the Softmax function to generate the final local attention weights. Local geometric features pass and Hadamard product operation ( This is obtained. Finally, the local feature offset is... Depend on The calculation yielded the result.

[0043] The outputs of the two branches then... With Δ The elements are added one by one and merged into the total feature offset. The total offset is then superimposed onto the original input feature X via residual connections to generate a new feature with depth correction. .

[0044] S430: Perform multi-scale feature fusion and classification.

[0045] Because the defect detection model has a hierarchical structure, the feature maps extracted from different levels first undergo a global max pooling operation, which then... The dimensions are aggregated into The feature vectors of all scales are then concatenated along the feature dimension to form a fused feature vector containing multi-scale information. This fused feature vector is finally fed into a classification head composed of a multilayer perceptron (MLP). The classification head outputs the final defect category probability distribution through a series of nonlinear transformations.

[0046] S450: Performs backpropagation and parameter updates.

[0047] During training, the cross-entropy loss function is used to quantify the difference between the model's predicted probability distribution and the true one-hot encoded labels. Based on the calculated loss value, the gradient of the loss with respect to all learnable parameters in the network is calculated using the backpropagation algorithm. The Adam optimizer updates the network parameters based on these gradients. Preferably, the hyperparameters of the Adam optimizer... and The values ​​are set to 0.9 and 0.999 respectively (both dimensionless), the initial learning rate is set to 0.001 (dimensionless), and a cosine annealing strategy is used to decay the learning rate; the training batch size is set to 16 (unit: samples / batch), and the total training cycles are set to 250, which is not limited here.

[0048] Example 2 Unlike Embodiment 1, this application transforms point clouds into a geometric representation rich in topological information and uses a graph neural network (GNN) for defect detection. This embodiment models the adjacency relationships between points, thereby enabling the model to more effectively capture complex defects defined by the relationships between points, such as the continuity of cracks or abnormal curvature changes on the surface.

[0049] Steps S100-S300 in this embodiment are the same as in Embodiment 1, and are intended to obtain a labeled, standardized point cloud dataset.

[0050] S500: Construct geometric structure data.

[0051] Before being input into the defect detection model, the graph construction module is activated. This module is configured to transform each standardized point cloud sample into a geometric graph. Specifically: S510: Define the vertices and edges of a graph.

[0052] The vertex set V of the graph is directly composed of N (e.g., 2048) data points in the point cloud; the edge set E of the graph is constructed based on the spatial proximity relationships between data points using the k-nearest neighbor (k-NN) algorithm; for each vertex... The algorithm searches for the k closest point in the point cloud in terms of spatial distance. (dimensionless) points, and the intersection of these points with... Undirected edges are established between the points to construct a graph that reflects the local topology of the point cloud.

[0053] S520: Calculate the initial node features and edge features.

[0054] To provide rich input information to the graph neural network, the graph construction module further calculates the initial geometric features for the vertices and edges of the graph.

[0055] For each vertex Its initial node feature vector Constructed as a 6-dimensional vector The vector is formed by concatenating the three-dimensional coordinates of the point and the local tangent plane normal vector fitted by principal component analysis (PCA) based on its k-NN neighborhood.

[0056] For connecting vertices and edge Its initial edge feature vector Constructed as a 4-dimensional vector This vector is formed by concatenating the relative position vectors between two points and the Euclidean distance.

[0057] S600: Feature extraction is performed using a graph neural network model.

[0058] The generated geometric graph with initial features is loaded into a graph neural network model centered on a Graph Attention Network (GAT). This model updates node features through an iterative message-passing mechanism. Specifically: In the l-th layer network, the features of node i Updated to The process includes the following operations: S610: Calculate the attention coefficient.

[0059] For node i and any of its neighboring nodes j, the attention coefficient It can be done The calculations show that W and a are learnable weight parameters.

[0060] S620: Normalized attention weights.

[0061] based on The final attention weights are generated by normalizing the values ​​across all neighbors of node i using the Softmax function. This weight determines the importance of neighbor j's information for node i's updates.

[0062] S630: Aggregate neighbor features.

[0063] New features of node i The features after transformation of all its neighboring nodes j To carry out Generated for the weighted summation of weights: Here, by stacking multiple layers of GAT, the final features of each node are... It can encode its multi-hop neighborhood structure information in the graph.

[0064] S700: Perform graph pooling and classification.

[0065] After multi-layer GAT processing, in order to obtain a single feature vector representing the entire heat dissipation substrate, a global average pooling layer is called to process the final feature vectors of all N nodes in the graph. Perform element-wise averaging to generate a graph-level feature vector. ,Should The data is then fed into an MLP classification head, which outputs the predicted probabilities of various defects. The training process of other models, including loss function calculation and parameter update, is consistent with S440 and S450 in Example 1, and will not be described in detail here.

[0066] The following examples illustrate this.

[0067] In one specific embodiment, the following steps are included: Step 1: Obtain 3D point cloud data of the heat dissipation substrate.

[0068] Industrial heat dissipation substrates on the production line were scanned from multiple angles using a binocular laser scanner to obtain raw 3D point cloud data containing spatial topographic information of the substrate surface. Ultimately, point clouds of 6398 heat dissipation substrate samples were obtained.

[0069] Step 2: Perform data preprocessing on the above dataset.

[0070] Since raw point cloud data typically contains background noise, outliers, and uneven point density, the following preprocessing operations are required: A radius filtering algorithm is used to remove detached noise points and outliers generated during the acquisition process. A farthest-point sampling algorithm is used to uniformly sample point clouds from different substrates to 2048 points to meet the fixed input dimension requirement of the neural network. Finally, the point cloud is normalized so that its coordinates lie within a unit sphere. This eliminates differences in spatial location and absolute scale between different substrates, allowing the model to focus more on learning the relative geometric features within them.

[0071] Step 3: Construct the training data set and the test dataset.

[0072] The preprocessed point cloud data is classified and labeled, and each substrate point cloud is assigned a real defect category label. In particular, to address the problem of imbalance between normal and defect samples in industrial scenarios, data augmentation operations (such as random rotation, translation, random scaling, or adding Gaussian noise) are performed on minority class samples to expand data diversity. The labeled and augmented point cloud dataset is randomly divided into training dataset, validation dataset, and test dataset according to a set ratio.

[0073] Step 4: Input the preprocessed training dataset into the model for training: The structure of the model is as follows Figure 2 As shown, the model is based on a hierarchical backbone network to enhance feature extraction, long-range contextual information modeling, and learning capabilities. Its core is a hierarchical backbone network composed of multiple "Sampling Group-Feature Transform (SG-Trans)" modules stacked together, and finally connected to a classification head for multi-scale feature fusion. The hierarchical backbone network first maps the point cloud into a series of multi-scale feature spaces through progressive downsampling and local feature aggregation. At each scale, our proposed dual offset attention module serves as the core transformation unit to deeply refine the features. Finally, the multi-scale features are fused and fed into the classification head to obtain the final classification result.

[0074] The dual offset attention module employs a residual offset learning mechanism to dynamically capture complementary biases between the global context and local geometry. Through a dual-track parallel architecture, the module is able to extract and fuse these complementary biases.

[0075] Global Offset Generator: This branch is responsible for capturing long-range, global contextual changes. We borrow the offset attention idea from PCT and employ an efficient self-attention mechanism. First, the input features are transformed through three independent linear transformations. Mapped to a Query, Key, Value matrix: To improve efficiency, the key and value matrices undergo an average pooling operation. Subsequently, self-attention features are calculated. Finally, the core of this branch is to calculate the difference between the input and the self-attention output, and process it through an LBR (Linear-BatchNorm-ReLU) network to obtain the global offset. .

[0076] Local offset generator: Used to capture local geometric deviations. Similar to the global offset generator, the input features... To enable local receptive fields for both the query and value, depthwise separable convolution is employed. This process yields an image containing local context information. , .pass With the global Hadamard Product The initial attention map is computed using a 1000-degree-of-freedom (FC) layer. It is then fed into a nonlinear gating network consisting of an FC layer, a Swish activation function, and a Tanh layer. This is to generate more expressive attention weights. The process can be described as follows: .in It is a learnable temperature coefficient. The generated attention weights are applied to the localization. The final local features are obtained from the above. Finally, calculate the local offset. .

[0077] Finally, the results calculated from the two parallel branches are... and The elements are added together and then fused using a 1x1 convolutional layer to obtain the total offset. This offset is applied to the original input via a residual connection. , It will go through a depthwise separable convolution to further enhance the interaction of local information during feature transformation.

[0078] Training process: After the point cloud is input into the network, during the feature transformation stage, the feature data is fed into the dual offset attention module. The global branch of this module calculates and extracts global feature offsets for long-range context, while the local branch combines depthwise separable convolution to extract local feature offsets for microscopic deformations. Subsequently, the two offsets are fused and superimposed onto the original features in residual form, completing the depth correction and refinement of the point cloud features. Multi-scale features extracted from different network levels are globally pooled and concatenated, then input into the classification head (Multilayer Perceptron, MLP) to output the predicted probability distribution of the substrate point cloud belonging to various defects. The loss value between the network's output predicted probability distribution and the true label is calculated (using the cross-entropy loss function or a combination of the invention's specific geometric perception loss function). Using an optimization algorithm (SGD optimizer), the gradient is calculated through backpropagation, and the weight parameters in the network are updated layer by layer.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0080] This embodiment also provides a deep learning-based defect detection system for heat dissipation substrates. This system is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0081] This invention provides a deep learning-based defect detection system for heat dissipation substrates, the system comprising: The data acquisition module is used to acquire the three-dimensional point cloud data of the heat dissipation substrate to be tested; The data preprocessing module is used to preprocess the three-dimensional point cloud data to generate a standardized input point cloud. The defect detection module is used to extract and fuse the geometric features of the input point cloud at multiple scales using a pre-trained defect detection model to obtain defect category prediction results.

[0082] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0083] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to perform the steps in any of the above method embodiments when executed.

[0084] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0085] Embodiments of the present invention also provide an electronic device including a memory and a processor, the memory storing a computer program and the processor being configured to run the computer program to perform the steps in any of the above method embodiments.

[0086] In one exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0088] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another apparatus, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0089] The units described as separate components may or may not be physically separate. A component shown as a unit can be one or more physical units; that is, it can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0091] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0092] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting defects in a heat dissipation substrate based on deep learning, characterized in that, include: Acquire the three-dimensional point cloud data of the heat dissipation substrate to be tested; The three-dimensional point cloud data is preprocessed to generate a standardized input point cloud; The geometric features of the input point cloud are extracted and fused at multiple scales using a pre-trained defect detection model to obtain defect category prediction results.

2. The method according to claim 1, characterized in that, The defect detection model performs in-depth correction and refinement of the features of the input point cloud through a dual offset attention module. The dual offset attention module includes a global offset generation branch and a local offset generation branch. The global offset generation branch is used to calculate the global feature offset representing the global context bias, and the local offset generation branch is used to calculate the local feature offset representing the local geometric distortion.

3. The method according to claim 2, characterized in that, The global offset generation branch generates global context features based on self-attention calculation of the input point cloud features, and determines the global feature offset based on the difference between the input point cloud features and the global context features. The local offset generation branch processes the input point cloud features through depthwise separable convolution to give them a local receptive field, calculates local geometric features based on the processed features, and then determines the local feature offset based on the difference between the input point cloud features and the local geometric features.

4. The method according to claim 1, characterized in that, Before inputting the standardized input point cloud into the defect detection model, the method further includes: The standardized input point cloud is transformed into geometric graph structure data; wherein, the data points in the input point cloud constitute the vertices of the geometric graph structure, and the edges between the vertices are determined based on the spatial proximity relationship between the data points.

5. The method according to claim 4, characterized in that, The defect detection model is a graph neural network model; The graph neural network model processes the geometric graph structure data through a message passing mechanism. Each vertex aggregates the feature information of its neighboring vertices to iteratively update its own feature representation, thereby enabling the extraction of local and global defect features.

6. The method according to claim 4, characterized in that, The method further includes: Calculate initial node features for each vertex of the geometric graph structure, the initial node features including the three-dimensional coordinates of the data point and the normal vector or curvature calculated based on its local neighborhood; Calculate initial edge features for each edge of the geometric graph structure, the initial edge features including the relative position vector or Euclidean distance between the two vertices connected by the edge.

7. The method according to claim 4, characterized in that, Preprocessing of the 3D point cloud data includes: A radius filtering algorithm is used to remove background noise points and outliers from the 3D point cloud data; The farthest point sampling algorithm is used to uniformly sample the three-dimensional point cloud data to a preset number of points; The sampled 3D point cloud data is normalized so that its coordinates lie within a unit sphere.

8. A deep learning-based defect detection system for heat dissipation substrates, characterized in that, include: The data acquisition module is used to acquire the three-dimensional point cloud data of the heat dissipation substrate to be tested; The data preprocessing module is used to preprocess the three-dimensional point cloud data to generate a standardized input point cloud. The defect detection module is used to extract and fuse the geometric features of the input point cloud at multiple scales using a pre-trained defect detection model to obtain defect category prediction results.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program is configured to perform the method described in any one of claims 1 to 7 when executed.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method as described in any one of claims 1 to 7.

Images