Wafer probe station inspection path planning method, apparatus, equipment and probe station

Abstract

This invention relates to the field of probe inspection technology, and discloses a method, apparatus, device, and probe station for wafer probe station inspection path planning. The method involves: performing boundary enhancement processing on a wafer surface image to obtain a boundary enhancement input tensor; performing dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map containing chip boundaries and test point positions; performing channel attention and spatial attention filtering on the first feature map to obtain a second feature map; performing chip distribution feature analysis on the second feature map to obtain test point clustering results; performing probe movement path planning to obtain a globally optimal path; establishing a probe station inverse kinematics model based on the globally optimal path, and calculating the probe trajectory control sequence based on the probe station inverse kinematics model. This invention enhances the ability to extract chip boundary features from the wafer surface, ensures the stability and smoothness of the probe during high-speed testing, and significantly improves probe positioning accuracy.

Description

Technical Field

[0001] This invention relates to the field of probe inspection technology, and in particular to a method, apparatus, equipment and probe station for planning inspection paths on a wafer probe station. Background Technology

[0002] Traditional wafer probe station inspection methods mainly employ serpentine and spiral scanning methods to plan the probe movement path. These methods do not fully consider the actual distribution characteristics of chips on the wafer surface, often resulting in excessively long probe movement distances and low testing efficiency. Furthermore, with the increase in wafer size and the increasing complexity of chip layouts, the problem of insufficient chip boundary recognition accuracy has become increasingly prominent, leading to probe positioning deviations and affecting testing accuracy.

[0003] Existing technologies do not pay enough attention to path planning in high-density chip areas, and cannot implement differentiated testing strategies for areas of different densities, resulting in unreasonable allocation of testing time and wasting valuable testing resources. In addition, traditional path planning algorithms are difficult to adapt to the high nonlinearity and uncertainty of chip layout on the wafer surface, lack in-depth analysis of the distribution of chip test points, and fail to perform intelligent path optimization based on chip layout characteristics. Summary of the Invention

[0004] The main objective of this invention is to provide a method, apparatus, device, and probe station for wafer probe station detection path planning. This invention enhances the ability to extract chip boundary features on the wafer surface, ensures the stability and smoothness of the probe during high-speed testing, and significantly improves the probe positioning accuracy.

[0005] To achieve the above objectives, the present invention provides a wafer probe station detection path planning method, comprising the following steps:

[0006] Boundary enhancement processing is performed on the wafer surface image to obtain the boundary enhancement input tensor;

[0007] Dual-path feature extraction is performed based on the boundary enhancement input tensor to obtain a first feature map containing the chip boundary and test point locations;

[0008] The first feature map is filtered using channel attention and spatial attention to obtain the second feature map;

[0009] Chip distribution feature analysis is performed on the second feature map to obtain the clustering results of the test points;

[0010] Based on the clustering results of the test points, probe movement path planning is performed to obtain the globally optimal path;

[0011] A probe station inverse kinematics model is established based on the global optimal path, and the probe trajectory control sequence is calculated based on the probe station inverse kinematics model.

[0012] The present invention also provides a wafer probe station detection path planning device, comprising:

[0013] The boundary enhancement module is used to perform boundary enhancement processing on the wafer surface image to obtain the boundary enhancement input tensor;

[0014] The feature extraction module is used to perform dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map containing the chip boundary and the test point location;

[0015] An attention filtering module is used to perform channel attention and spatial attention filtering on the first feature map to obtain a second feature map;

[0016] The distribution feature analysis module is used to perform chip distribution feature analysis on the second feature map to obtain the test point clustering results;

[0017] The path planning module is used to perform probe movement path planning based on the clustering results of the test points to obtain the globally optimal path;

[0018] The calculation module is used to establish a probe station inverse kinematics model based on the global optimal path, and to calculate the probe trajectory control sequence according to the probe station inverse kinematics model.

[0019] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0020] The present invention also provides a probe station for implementing the steps of the method described in any of the above embodiments.

[0021] In summary, this invention significantly improves chip boundary recognition accuracy by introducing dual-path supervision and attention filtering techniques, enabling the wafer probe station to more accurately locate test points and effectively solving the problem of inaccurate chip boundary recognition in existing technologies. The method combining multi-scale convolutional neural networks and boundary extraction operators enhances the ability to extract chip boundary features from the wafer surface, making probe positioning more accurate and reducing the test error rate. Based on the feature analysis method using the chip distribution covariance matrix and spectral clustering algorithm, a deep understanding of the chip distribution characteristics on the wafer surface is achieved, enabling reasonable test area division based on chip density characteristics. Through density-level region division and priority weight allocation strategies, this invention implements differentiated testing schemes for different density regions, optimizing resource allocation and greatly improving testing efficiency. A hybrid algorithm combining multi-objective particle swarm optimization and adaptive annealing-taboo search is proposed, achieving global optimization of probe paths on complex wafer surfaces, significantly reducing probe movement distance and testing time. By establishing an accurate probe station inverse kinematics model and combining it with variable step-size B-spline curve interpolation for trajectory planning, the stability and smoothness of the probe during high-speed testing are ensured, significantly improving probe positioning accuracy. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the steps of a wafer probe station detection path planning method in one embodiment of the present invention;

[0023] Figure 2 This is a structural block diagram of a wafer probe station detection path planning device according to an embodiment of the present invention;

[0024] Figure 3 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0025] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0027] Reference Figure 1 This embodiment provides a wafer probe station detection path planning method, including the following steps:

[0028] S1, perform boundary enhancement processing on the wafer surface image to obtain the boundary enhancement input tensor;

[0029] In this process, wafer surface images are input into three convolutional layers with kernels of different sizes for feature extraction. These three convolutional layers use standard convolutional kernels of 3×3, 5×5, and 7×7 sizes, respectively, to perform parallel feature extraction. While preserving the original image structure, this captures the spatial variation features of the wafer chip edges at different scales. Each convolutional layer has a uniform stride parameter of 1 and a fixed number of output channels of 32 to ensure consistent channel dimensions across different scales, facilitating subsequent feature fusion. A feature fusion module performs a weighted fusion operation on the three feature maps. This module uses a bitwise weighted summation method for fusion, where the weight coefficients corresponding to the outputs of the three convolutional kernels are initialized to 0.4, 0.35, and 0.25, respectively. During fusion, the weight coefficients are dynamically adjusted through the backpropagation mechanism of the neural network, allowing the model to adaptively select the most representative scale feature information. After fusion, a fused feature map integrating multi-scale edge perception capabilities is obtained. Based on the fused feature map, boundary extraction operators are introduced to perform edge detection operations. These include a set of Sobel operators for extracting horizontal edges, a set of Sobel operators for extracting vertical edges, and a set of Laplacian operators for omnidirectional edge extraction. The outputs of these three edge operators are weighted and fused according to a weight ratio of 1:1:0.8 to generate a boundary feature map. An adaptive threshold segmentation algorithm is applied to the boundary feature map to dynamically calculate a threshold T suitable for the current image feature distribution. The calculation formula is T=μ+k×σ, where μ is the mean of the boundary feature map, σ is the standard deviation, and k is an adjustment coefficient, initially set to 1.5. This threshold is used to perform a binarization operation on the boundary feature map, effectively separating salient edge regions from the background and generating a boundary enhancement map. The boundary enhancement map is then concatenated with the original wafer surface image along the channel dimension to form a multi-channel input tensor, which constitutes the boundary enhancement input tensor for subsequent deep feature extraction and path planning analysis.

[0030] S2, perform dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map containing the chip boundary and test point location;

[0031] Specifically, the boundary enhancement input tensor is input into the encoder part of the main path structure. This encoder, based on an improved U-Net architecture, contains four downsampling modules. Each module consists of two 3×3 convolutional layers, one batch normalization layer, and one max pooling layer, effectively compressing spatial dimensions and extracting semantic information at different levels. Through layer-by-layer downsampling, multi-level feature representations from local edge textures to global structural contours are gradually captured, outputting four encoded feature maps at different spatial scales. These feature maps correspond to the appearance of the wafer image under different receptive fields, exhibiting strong hierarchicality and spatial distribution recognition capabilities. The four encoded feature maps are then decoded layer by layer by the main path decoder. The decoder also contains four upsampling modules, each including a transposed convolutional layer and two 3×3 standard convolutional layers, along with a batch normalization layer, to restore the spatial dimension of the image and fuse multi-scale semantic information. During the decoding process, each layer is skipped to the encoded feature map of the corresponding scale downsampling stage to enhance edge and structure restoration capabilities, reconstructing a high-resolution main path decoded feature map. While the main path completes the decoding process, three representative feature maps from different levels are selected from the four feature maps output from the main path encoder stage as input sources for the supervision path. These maps are then fed into three independent branches of the supervision path. Each supervision branch contains a 1×1 convolutional layer for channel dimensionality reduction, a 3×3 convolutional layer for local feature extraction, and an upsampling layer for size alignment. This ensures that feature maps from different depths can be uniformly mapped to the features decoded in the main path. Figure 1 The spatial scale is optimized. Dimensionality reduction effectively reduces computational complexity while allowing shallow detail features and deep semantic features to be extracted and enhanced in a lightweight manner. The feature maps of the three supervised branches are fused with the decoded feature map of the main path through a size adjustment operation. The fusion process is completed using feature concatenation or bit-weighted summation to construct a fused feature representation. This fused feature representation is then input into the semantic segmentation module for processing. The semantic segmentation module uses a fully convolutional structure for pixel-by-pixel classification. Based on the structural boundary cues and test point response information in the fused features, a high-resolution feature map containing the wafer chip boundary and test point locations is output, which is the first feature map.

[0032] S3, perform channel attention and spatial attention filtering on the first feature map to obtain the second feature map;

[0033] It should be noted that the first feature map undergoes saliency modeling operations along the channel dimension, specifically global average pooling and global max pooling. These two operations compress the data along the spatial dimension, preserving the global statistical features of each channel. Average pooling reflects the overall response intensity of the channel, while max pooling emphasizes the response peak of the most salient region. Together, they form a statistical vector describing the importance of the channel. These two statistical vectors are then input into a network structure consisting of two fully connected layers. Dimensionality reduction (1 / 16 of the channel number) is performed using FC1, followed by ReLU activation and then input into FC2 for dimensionality upscaling, restoring the vector to the original channel number to form a finely expressed channel attention weight vector. The outputs of the two channel paths are summed along the channel dimension and normalized using the Sigmoid function to obtain channel weight coefficients ranging from 0 to 1, used to characterize the importance of each channel. Average pooling and max pooling are performed on the first feature map along the channel dimension to generate two spatial single-channel feature maps. The former focuses on the average response region of all channels, while the latter emphasizes the spatial distribution information dominated by the maximum value in the channel. These two methods capture the features of spatially salient regions from the perspectives of balance and extrema, respectively. The two single-channel images are stitched together along the channel direction to form a two-dimensional structure map, which is then input into a 7×7 convolution kernel for convolution to capture local spatial relationships within a larger receptive field. The convolution result is activated by the Sigmoid function to generate a spatial attention weight map with the same spatial dimension as the original image, identifying regions that need to be preserved and suppressed at the spatial level. The channel attention weight vector is multiplied with the first feature map channel by channel, adjusting the response intensity of each channel and retaining information-rich and task-relevant feature channels. Subsequently, this result is multiplied with the spatial attention weight map pixel by pixel, enhancing the response capability of salient regions at the spatial level and weakening interference from useless regions. This combined attention mechanism completes dual saliency screening in both the channel and spatial dimensions, outputting a preliminary filtered feature map that integrates channel and spatial screening. Based on the position of the preliminary filtered feature map in the network, residual attention mechanisms with different coefficients are applied to the shallow and deep feature maps respectively, dynamically adjusting the screening intensity according to the position of the feature map in the network. For shallow feature maps near the input, which focus on capturing local edges and chip contour information, the original structural details are preserved, and conservative feature enhancement is achieved through a low-intensity residual enhancement strategy. For deep feature maps near the output, however, the focus is on expressing higher semantic targets such as test points, employing a stronger residual enhancement strategy to significantly improve the expression of high-level abstract features. By applying differentiated enhancements to the shallow and deep feature maps respectively, the information flow is progressively strengthened and focused from low-level details to high-level semantics throughout the network structure, resulting in the second feature map.

[0034] S4. Perform chip distribution feature analysis on the second feature map to obtain the clustering results of the test points;

[0035] Specifically, the second feature map is input into a semantic segmentation network, which possesses pixel-level region recognition capabilities. This network can accurately segment various semantic regions in the wafer image and output a binary distribution map containing only test point regions. In this image, all pixels identified as test points have a value of 1, and the rest have a value of 0. This operation defines the spatial distribution of test points on the wafer surface. After semantic segmentation, connectivity analysis is performed on the test point distribution map to identify each connected cluster of test points. Representative basic feature information is extracted from each cluster, including the center position of the region, the area occupied, and the number of test points within the region. The center coordinates reflect the geometric position of the chip or region on the entire wafer, the area describes its spatial coverage, and the number of test points is an important parameter for measuring the detection complexity of the region. After integrating these data, a basic description of multiple independent chip regions is obtained. Based on this, the spatial distribution structure of the chip is analyzed by comparing the positional relationships and attribute characteristics between different regions. By constructing a numerical representation of the similarity between regions, the degree of difference between each pair of chip regions is quantified, resulting in the chip distribution covariance matrix. A spectral clustering algorithm is applied to the chip distribution covariance matrix, where the number of clusters is adaptively determined based on the relative density between regions, thus dividing the entire wafer surface into several groups of structurally and functionally similar regions. The test point density within each cluster is statistically analyzed; test point density refers to the number of test points contained in a unit area of ​​a chip cluster. Regions with higher density have greater detection complexity and more concentrated testing time, and are therefore considered high-priority testing regions. By ranking the test densities of all clusters, a test priority classification standard is established, ensuring that probes prioritize processing the most complex and densest test regions during actual movement, improving overall testing efficiency. For high-priority regions, more refined test point clustering is performed. A clustering algorithm sensitive to locally dense structures is selected to further subdivide the test points within each high-density region, forming multiple smaller-granularity, more compact test point groups, ultimately yielding test point clustering results with clear boundaries, reasonable structure, and test priority.

[0036] S5. Based on the clustering results of the test points, perform probe movement path planning to obtain the globally optimal path;

[0037] Based on the clustering results of the test points, the density distribution features and spatial distribution features of each cluster region are extracted. The density distribution features reflect the density of test points in each region, while the spatial distribution features describe the arrangement and geometric expansion characteristics of the test points within a planar range. This information is uniformly encoded into several descriptive vectors, forming a multi-dimensional feature vector matrix. Nonlinear weighted calculations are performed on the feature vector matrix to establish a path planning objective function that comprehensively considers the influence of multiple factors. This function integrates multiple indicators such as travel distance, test order, regional complexity, and path coherence, forming an optimization evaluation criterion that reflects the balance between probe movement efficiency and test coverage efficiency. The clustering results of the test points are divided into multiple density level regions, and different test weights are configured according to the detection priority represented by each level. The weights determine the priority processing order and access frequency of high-density regions in probe path planning, thereby constructing a hierarchical detection sequence system. Based on this hierarchical structure, path construction is performed individually for each test point within each density level region. Using heuristic or fast iterative shortest path search strategies, corresponding local sub-paths are generated for different regions. The optimal path set within each region is selected based on criteria such as path compactness, reasonable inter-point connections, and local testing cost. This set represents the optimal test path structure within a fixed region. After completing the local path construction, the dispersed optimal sub-paths are combined into a global path system, and a multi-objective particle swarm optimization algorithm is introduced for global connectivity planning. This optimization process aims to minimize the connection cost between regions, the overall path continuity, and path overlap. It continuously iterates on the path connection schemes between regions, dynamically adjusting particle positions and state evolution to find the optimal connection strategy, forming a preliminary global path structure that orderly connects the various locally optimal paths. A hybrid optimization strategy combining adaptive annealing and tabu search is applied to the initial connection path. Simulated annealing introduces a probabilistic escape mechanism to avoid getting trapped in local minima. Simultaneously, combining historical state records from tabu search with path replacement strategies effectively suppresses path structure from getting stuck in repeated updates and cyclic redundancy. This allows for iterative optimization of the probe movement path and efficient fine-tuning of local details at a global scale, resulting in the globally optimal path.

[0038] S6. Establish the probe station inverse kinematics model based on the global optimal path, and calculate the probe trajectory control sequence based on the probe station inverse kinematics model.

[0039] Specifically, a three-dimensional coordinate sequence of test points is extracted from the globally optimal path. This sequence describes the spatial positions of the probe on the wafer surface as it sequentially visits the planned path. The three-dimensional coordinate data is paired one-to-one with the three-axis control parameters of the actual probe station to construct a motion mapping sample set containing the actual control response. The control parameters include the probe station's horizontal movement angle, precise vertical displacement, and rotation angle to ensure attitude consistency. Feature standardization and spatial transformation are performed on the probe station motion mapping samples. By unifying the scale range, the coordinate data and control parameters in each sample are transformed to a normalized numerical range. Rotation alignment and coordinate system mapping are performed while maintaining the geometric structure, thereby constructing a normalized feature vector set with a unified reference frame and numerical computability. Based on the normalized feature vectors, a multi-parameter nonlinear regression equation system describing the probe station's control behavior is constructed. This equation system contains three sets of independent but mathematically related nonlinear functions. Each set of equations corresponds to the control commands on the three degrees of freedom of the probe station: the calculated expression of the horizontal angle, vertical displacement, and rotation angle. To improve solution accuracy and model convergence performance, the regression equations are embedded in a hybrid solver constructed from particle swarm optimization (PSO) and neural networks. Leveraging the global search capability of PSO in continuous space and the local accuracy advantage of neural networks in function fitting, the coefficients of the regression equations and the weight parameters of the neural network are iteratively updated alternately. This ensures that the entire inverse kinematics model possesses both the ability to accurately model complex motion relationships and good convergence and adaptive performance. After model construction, the spatial position difference between adjacent test points on the globally optimal path is used as input and sequentially fed into the inverse kinematics model to calculate the corresponding probe displacement control sequence. Each control quantity describes the single-step horizontal angle change, longitudinal displacement step size, and rotation adjustment step size required to move from the current position to the next test point. These control quantities are combined item by item to form a continuous sequence of control commands, constituting the control framework for the probe throughout its entire journey from the starting point to the end point. To improve the stability of probe operation and the continuity of mechanical response, a curve fitting and trajectory smoothing mechanism is introduced based on the displacement control sequence. Through cubic spline interpolation or high-order Bezier fitting algorithms, curvature optimization and redundancy suppression are performed on the connection segments between control quantities, making the velocity, acceleration, and angle changes of the probe during execution smoother, thereby effectively avoiding system oscillation, error accumulation, and mechanical fatigue. The probe trajectory control sequence is output through the above process.

[0040] In one example, boundary enhancement processing is performed on a wafer surface image to obtain a boundary enhancement input tensor, which includes:

[0041] The wafer surface image is input into three convolutional layers with convolutional kernels of different sizes for feature extraction, resulting in three sets of feature maps;

[0042] The three sets of feature maps are weighted and fused to obtain a fused feature map;

[0043] Edge detection is performed on the fused feature map using boundary extraction operators to obtain the boundary feature map. The boundary extraction operators include the horizontal Sobel operator, the vertical Sobel operator, and the Laplacian operator.

[0044] Adaptive thresholding is performed on the boundary feature map to obtain a boundary enhancement map. The boundary enhancement map is then concatenated with the wafer surface image to obtain the boundary enhancement input tensor.

[0045] In this example, the raw wafer surface image is input into three convolutional layers with different receptive fields, resulting in three sets of feature maps. These three convolutional layers use kernels of sizes 3×3, 5×5, and 7×7, respectively, to process the input image, capturing structural information and boundary details in the wafer image at different scales. Smaller kernels are more sensitive to local texture changes, while larger kernels help capture broader background contours and shape structures. The three feature maps are then weighted and fused using a weighted fusion mechanism. Different fusion weights are set for each feature map to dynamically adjust the proportion of features at different scales in the overall representation. Initial weights are pre-set and then continuously optimized during network training through backpropagation, allowing the model to autonomously learn which type of feature contributes more to boundary recognition and achieve optimal feature fusion. The fusion method uses a pixel-wise weighted approach, summing the feature values ​​at corresponding positions in the three channels according to the set weights. This preserves the response advantages of multi-scale features in the fused image while eliminating redundant information and inefficient activations, generating a fused feature map. Edge detection is performed on the fused feature map using boundary extraction operators, including the horizontal Sobel operator, the vertical Sobel operator, and the Laplacian operator. The Sobel operator detects edges by calculating the image's gray-level gradient. The horizontal Sobel operator detects lateral boundary changes in the image, while the vertical Sobel operator extracts vertical edge contours; together, they outline the main boundary direction changes. The Laplacian operator, a second-derivative operator, has omnidirectional detection capabilities, responding to all edge changes in the image and excelling at capturing structural abrupt changes and high-frequency boundary signals. In practical applications, the fused feature map is input into each of the three edge detection operators for independent calculation, resulting in three edge response maps for corresponding directions. These three maps are then weighted and fused. The fusion weights are set or optimized based on the strength distribution of edge responses in the actual scene to ensure that the final boundary feature map retains the true chip boundaries to the greatest extent while suppressing noise and false edges. The boundary feature map is binarized, transforming the edge map of continuous pixel values ​​into a binary image containing only edges and non-edges. An adaptive thresholding method is employed, dynamically generating the optimal segmentation threshold based on the image's statistical characteristics to adapt to variations in brightness, contrast, and edge strength across different wafer images, avoiding over-segmentation or under-segmentation issues caused by using a fixed threshold. During calculation, the average pixel value of the entire boundary map is selected as the base, and the final threshold is dynamically determined by combining the pixel standard deviation and adjustment parameters. This threshold is used to classify pixel values ​​in the boundary feature map; values ​​above the threshold are classified as boundaries, and values ​​below the threshold are classified as non-boundaries, resulting in a boundary enhancement map with a stable boundary contour structure.The boundary enhancement map is stitched together with the original wafer image at the channel level, and different information dimensions are input into the neural network in a parallel structure to form a multi-channel boundary enhancement input tensor.

[0046] In one example, dual-path feature extraction is performed based on the boundary enhancement input tensor to obtain a first feature map containing the chip boundary and test point locations, including:

[0047] The boundary enhancement input tensor is fed into the encoder of the main path for multi-level feature extraction, resulting in four encoded feature maps at different scales. The encoder contains four downsampling modules.

[0048] The decoder of the main path decodes the coded feature maps at four different scales to obtain the main path decoded feature map. The decoder contains four upsampling modules.

[0049] Three feature maps of different levels are selected from the main path encoding feature map and input into the supervision path to obtain three supervision branch feature maps;

[0050] Channel dimensionality reduction and feature extraction are performed on the feature maps of the three supervised branches respectively to obtain supervised feature maps. The supervised feature maps are then sizing and fused with the main path decoding feature map to obtain the fused feature representation.

[0051] Semantic segmentation is performed on the fused feature representation to obtain a first feature map containing chip boundaries and test point locations.

[0052] In this example, the encoder structure that inputs the boundary enhancement input tensor into the main path is based on an improved U-Net architecture and consists of four consecutive downsampling modules. Each module includes two convolutional operations, a batch normalization layer, and a max pooling layer. The convolutional operations use small-sized kernels to extract local spatial features, the batch normalization layer accelerates convergence and stabilizes the training process, and the max pooling operation scales down the spatial dimension, thereby progressively reducing the resolution and increasing the receptive field of the features. During the four downsampling processes, the spatial size of the feature maps decreases sequentially, while the channel dimension gradually increases. The layer-by-layer extracted encoded feature maps retain structural semantic information at different scales. Shallower feature maps focus more on edge texture and spatial localization, while deeper layers form an abstract representation of the overall distribution of chip arrangement patterns and test point structures. This layer-by-layer extraction mechanism enables the model to comprehensively understand the spatial composition of the wafer image from local to global perspectives. After encoding, the encoded feature maps at four scales are sequentially fed into the decoder structure. This decoder consists of four upsampling modules, each including a transposed convolution operation, two standard convolutional layers, and a batch normalization layer. The transposed convolution operation restores the feature map size to the original image scale layer by layer, while the convolution operation fuses the current layer information with the interpolated features from the upsampling process. Furthermore, a skip connection mechanism directly introduces features from the encoder's corresponding layer into the decoder to improve edge information restoration and detail preservation. After four rounds of upsampling, the main path decoded feature map is obtained. Three representative and feature-discriminative outputs from the four feature maps generated in the main path encoding stage are selected and used as inputs to three independent branches of the supervised path. These three sets of feature maps represent shallow, mid-level, and deep features, respectively, and their spatial distribution and local contour information are complementary. The supervised path treats them as parallel processing objects, performing a series of operations in their respective branches to extract task-relevant feature representations. Each supervised branch undergoes channel dimensionality reduction through a one-to-one convolution operation. This operation reduces redundant information and compresses computational complexity, ensuring that features at different levels do not lose alignment due to channel inconsistencies during subsequent processing. After dimensionality reduction, standard convolution is performed to enhance feature discrimination capabilities. Finally, upsampling is used to adjust the supervised feature maps output from different levels to the same spatial size as the main path decoded feature map, providing an aligned input data structure for subsequent fusion processing. After size adjustment, these three supervised feature maps are fused with the main path decoded feature map. This fusion process employs either channel concatenation or weighted summation, selecting the more suitable fusion strategy based on the actual training situation. By fusing the fine-grained, multi-layered boundary structures in the supervised path with the global, coarse-to-fine semantic information in the main path, the resulting fused feature representation possesses good boundary clarity in spatial structure and combines macroscopic chip location understanding with microscopic observation pilot distribution control capabilities at the semantic level.The fused feature representation is input into the semantic segmentation module, which uses a fully convolutional structure for pixel-by-pixel classification and combines multi-level convolutional kernels with a context extraction mechanism to perform boundary judgment and target point identification for each pixel in the wafer image. During training, a joint optimization objective combining cross-entropy loss and region overlap loss is adopted, enabling the model to focus on both overall classification accuracy and boundary fitting accuracy in small-region test point extraction. The final output first feature map is a spatial dimension equal to the original. Figure 1 The image contains a structural diagram that depicts the outline of the wafer chip and identifies the specific locations of all test points within the image.

[0053] In one example, channel attention and spatial attention are applied to the first feature map to obtain the second feature map, which includes:

[0054] Perform global average pooling and global max pooling operations on the first feature map to obtain channel-dimensional statistical features;

[0055] The statistical features of the channel dimension are input into two fully connected layers for dimensionality reduction and dimensionality increase, respectively, to obtain the channel attention weight vector.

[0056] The first feature map is subjected to average pooling and max pooling along the channel dimension to obtain two single-channel feature maps.

[0057] The two single-channel feature maps are concatenated and then fed into a 7×7 convolutional layer to obtain the spatial attention weight map.

[0058] Multiply the channel attention weight vector by the first feature map by channel, and then multiply it element-wise by the spatial attention weight map to obtain the preliminary filtered feature map;

[0059] Based on the position of the initially selected feature maps in the network, residual attention mechanisms with different coefficients are applied to the shallow and deep feature maps respectively to obtain the second feature map. The residual attention mechanism is used to preserve chip layout details and test point distribution information.

[0060] In this example, the first feature map is compressed spatially. Global average pooling and global max pooling are used to statistically analyze the responses in each channel dimension into two one-dimensional vectors. Global average pooling reflects the overall response intensity trend within the channel, while global max pooling captures the distribution of the strongest response points within the channel. These two sets of channel statistical features are then input into two structurally identical fully connected networks. Each fully connected network consists of two cascaded linear transformation layers. The first layer performs dimensionality reduction, shrinking the original number of channels to a fixed proportion. The compressed intermediate representation is then processed by a non-linear activation function before being input into the second fully connected layer for dimensionality increase, restoring the original number of channels. Through this dimensionality reduction-increase structure, the model focuses on the correlation and dependency structure between each channel during learning, forming a set of attention weight vectors with channel-aware capabilities. The two channel attention weight vectors are element-wise summed and fused, and then compressed to a normalized range using an activation function, outputting a set of attention coefficients consistent with the original number of channels. Simultaneously, to enhance the model's ability to recognize spatial structures, a spatial attention branch is constructed. This branch takes the first feature map as input and performs aggregation operations on it along the channel dimension. Through average pooling and max pooling along the channel dimension, it generates two single-channel two-dimensional feature maps. These two maps reflect the mean distribution of the accumulated responses across all channels and the location structure of the extreme responses, respectively. They provide complementary perspectives on chip layout and structural contours in a spatial dimension. Subsequently, these two maps are concatenated along the channel dimension into a dual-channel intermediate tensor, which is then input into a 7×7 two-dimensional convolutional layer for local spatial relationship modeling. This convolutional operation has a large receptive field, capable of capturing microstructural features such as boundary lines, region contours, and spatial repetition within the neighborhood. After non-linear activation, the convolutional output yields a spatial attention weight map of the same size as the input map. This map indicates the attention intensity at each pixel location within the two-dimensional space. The channel attention weight vector is multiplied by the first feature map channel by channel to adjust the response intensity of each channel, achieving amplification of high-response channels and compression of low-response channels. The intermediate feature map, after channel reweighting, is then multiplied pixel-by-pixel with the spatial attention weight map to form a preliminary filtered feature map that integrates channel-dimensional selectivity and spatial-dimensional attention. To enhance the adaptability of the feature map to the task objective and suppress information weakening during network propagation, a differentiated residual attention mechanism is applied based on the current feature map's position within the overall network structure.For feature maps located in shallow networks, the main task is to preserve the clear details and well-defined boundaries of the chip outline in the wafer image. Therefore, a small residual enhancement strategy is adopted, where the original input feature map and the selected feature map are residually superimposed with a small coefficient to enhance the expression of salient regions while maintaining the original texture. For feature maps located in deep networks, which carry more semantic-level global expressions such as test point distribution, density state, and layout structure, a larger residual superposition strategy is adopted to amplify local anomalies, clustered high-density regions, and spatial center features shown in the selected feature map, thereby improving the model's ability to identify test regions. After being corrected by the corresponding residual coefficients, the shallow and deep feature maps are respectively used to generate complete second feature maps.

[0061] In one example, chip distribution feature analysis is performed on the second feature map to obtain the test point clustering results, including:

[0062] Semantic segmentation is performed on the second feature map to obtain a binarized test point distribution map;

[0063] Perform connectivity analysis on the binarized test point distribution map to obtain the basic features of the chip region and test points. The basic features include the center coordinates, area and number of test points.

[0064] A chip distribution covariance matrix is ​​constructed based on basic features, and a spectral clustering algorithm is applied to the chip distribution covariance matrix to obtain K chip clusters;

[0065] For each chip, cluster the test point density and sort them to obtain test priority region division information. The test point density is defined as the number of test points per unit area.

[0066] Based on the test priority region division information, fine-grained clustering is performed on the test points in high-density regions to obtain the test point clustering results.

[0067] In this example, the second feature map is input into the semantic segmentation module to identify regions of test significance in the wafer image and accurately label the locations of test points within them. The semantic segmentation network employs a fully convolutional architecture and a multi-scale feature fusion strategy to acquire high semantic information while maintaining spatial resolution. Through the network's pixel-level classification capability, a classification image with the same size as the input image is output. This output image is converted into a binary test point distribution map, where all pixels identified as test points are uniformly marked as 1, and other regions are marked as 0, forming a structure map representing the spatial distribution of test points. Connectivity analysis is performed on this binary map. This operation, based on the connectivity principle of adjacent pixels having the same value, identifies all independent regions composed of test point pixels. Each connected region corresponds to a physical chip unit or sub-test area. By traversing each connected region and extracting its basic geometric features, a basic data description of the chip distribution and test point clusters is obtained. The center coordinates are calculated using pixel centroids to indicate the region's position on the entire wafer plane; the area is determined by the number of pixels in the calculated region, reflecting the space occupied by the chip on the wafer; and the number of test points is obtained by counting the total number of pixels marked as test points within the region, revealing the density and complexity of the testing tasks within that region. These features collectively constitute the basic dataset for each chip region. After standardizing the basic feature vectors of all chip regions, a covariance matrix is ​​constructed. This matrix measures the comprehensive similarity between any two regions in terms of spatial location, area size, and test density. Based on this, a spectral clustering algorithm is introduced to process the covariance matrix, mapping the data points to a low-dimensional space where clustering is performed using the graph structure between the data. By performing eigenvalue decomposition on the matrix and constructing a graph Laplacian matrix, the model identifies regions distributed on the wafer with similar structures or test intensities and groups them into the same category, thus dividing the original chip distribution structure into K clusters, each representing a group of physically or logically closely related chip regions. The value of K is adaptively adjusted using unsupervised metrics such as the silhouette coefficient to achieve a balance between the number and quality of clusters. After obtaining these clustering results, the test point density of each chip cluster is calculated to determine the priority in path planning. The density is obtained by the ratio between the number of all test points in the cluster and the total area it covers. The higher the density, the more concentrated the test points and the more frequent the path switching, and therefore it is considered a high-complexity, high-priority test area. After sorting all clusters according to their test density from high to low, a set of test priority area division information is generated. This information is used in the path optimization and task scheduling stages to determine the order in which probes enter the regions first, thereby improving overall testing efficiency and reducing unnecessary movement.For regions with extremely high test density, the test points are densely distributed without a clear structural pattern. Therefore, finer-grained clustering is required for these high-density regions. Fine-grained clustering methods, such as density-based region clustering, are introduced to identify subsets of test points that are tightly clustered in local space, naturally relegating sparse or outlier points to the periphery or removing them. By performing fine-grained partitioning on the test points in high-density regions, the complex structure is transformed into several manageable sub-blocks, providing locally optimized spatial units for subsequent path planning. This allows the probe to form short, efficient sub-paths within each sub-block when performing test tasks, and then integrates these sub-paths into a unified test path.

[0068] In one example, probe movement path planning is performed based on the test point clustering results to obtain the globally optimal path, including:

[0069] Based on the clustering results of the test points, the density distribution features and spatial distribution features of the test points in each cluster are extracted to obtain the feature vector matrix;

[0070] A nonlinear weighted calculation is performed on the eigenvector matrix to obtain the path planning objective function;

[0071] The clustering results of the test points are divided into multiple density level regions, and a test priority weight is assigned to each density level region to obtain a hierarchical detection sequence;

[0072] Based on the hierarchical detection sequence, a local sub-path is constructed for the test points within each density level region to obtain the set of optimal paths within the region;

[0073] Multi-objective particle swarm optimization is performed on the optimal path set within the region to obtain the path connection scheme between regions;

[0074] Based on the inter-regional path connection scheme, adaptive annealing-tabo search hybrid optimization is performed on the global path to obtain the globally optimal path.

[0075] In this example, structural features are extracted for each cluster region to construct a feature vector matrix with numerical descriptive capabilities. For each test point cluster, the average distribution density of test points within a unit area is calculated, and its spatial layout features are extracted, including centroid coordinates, geometric boundary shape, principal direction expansion angle, and average, maximum, and minimum spacing between test points. These features together constitute a complete multi-dimensional vector, reflecting the complexity within the cluster and embodying the relative position and density characteristics of the region within the overall wafer structure. The feature vectors of all cluster regions are arranged row-wise to form a feature vector matrix. Weighted calculations are performed on different feature dimensions in the feature vector matrix. Since different feature dimensions have different physical meanings and nonlinear differences in their impact on the path structure, the weighting method needs to be set in conjunction with actual test constraints. A nonlinear mapping strategy is used to transform each feature value according to the cost it generates in the actual path, and then the transformed values ​​are added according to the set weight coefficients to generate a comprehensive evaluation index for path planning. This metric, serving as the objective function for path optimization, encompasses fundamental requirements such as path length, turning complexity, and test point coverage efficiency. It also comprehensively evaluates density variations, regional transition continuity, and control action switching frequency within the path, guiding the optimization algorithm towards convergence towards shorter overall movement distances, more rational test sequences, and higher mechanical execution efficiency. After constructing the objective function, the test point clustering results are restructured, dividing all clustered regions into multiple density levels. The division criteria are based on a joint judgment of test point density and spatial scale for each cluster, categorizing them into high-density, medium-density, and low-density areas. High-density areas correspond to regions with concentrated test tasks, imposing higher requirements on path control precision and time efficiency, while low-density areas appropriately relax local constraints on path planning. Based on the density level division, different test priority weights are assigned to different density regions; higher weights result in higher priority for the region in the test execution sequence, thus constructing a hierarchical detection sequence mechanism. Based on this hierarchical detection sequence, a local path is constructed for the test points within each density level region. Within each region, multiple feasible test point access paths are generated by combining all test points using distance heuristics, nearest neighbor strategies, or graph search algorithms. Then, locally optimal paths are selected based on metrics such as path length, angle variation, and jump distance between points. This operation ensures the shortest path connecting points to points within the region, while also considering the boundary smoothness and connectivity of this path when connecting to the global path. Therefore, a set of connectable edge points must be retained at the end of each local path for subsequent global path concatenation. After all local paths are constructed, they are input as sub-task units into the multi-objective particle swarm optimization module. Particle swarm optimization considers not only the connection order between paths but also the turning cost between path segments, connection jump distance, overall path smoothness, and path execution time.By globally evaluating the path sequence of each particle and continuously adjusting the particle decoding strategy and inter-particle information interaction mechanism during iteration, a path connection scheme that achieves an optimal balance between global shortest distance and region connectivity efficiency is gradually converged. This scheme is output as a framework for the final path structure, using a specific region connection order and connection point integration method. Based on the determined inter-region path connection scheme, a hybrid optimization mechanism combining adaptive annealing and tabu search is introduced to eliminate local discontinuities, redundant detours, and repeated node visits in the path. The adaptive annealing algorithm allows for the acceptance of inferior solutions to a certain extent to escape local optima, thereby improving the global optimal convergence probability of the overall path structure; while tabu search avoids getting stuck in repeated loops and invalid transformations by recording historical solution states and forbidden regions. The combination of the two simultaneously solves the problems of global optimization and local trap control. During the optimization process, local segments of the path will undergo perturbation transformations, such as node swapping, path segment reversal, and segment interpolation. All transformed results are re-evaluated based on the objective function and selected for retention according to probability rules. The final output path achieves overall optimization in multiple dimensions such as total path length, number of turns, sub-path connection stability, and execution coherence.

[0076] In one example, a probe station inverse kinematics model is established based on the globally optimal path, and the probe trajectory control sequence is calculated based on the probe station inverse kinematics model, including:

[0077] Extract the three-dimensional coordinate sequence of the test point from the global optimal path, and construct the probe station motion mapping sample by combining the three-dimensional coordinate sequence with the corresponding probe station three-axis control parameters. The probe station three-axis control parameters include the probe station's horizontal movement angle, vertical displacement, and rotation angle.

[0078] Perform feature standardization and spatial transformation on the probe station motion mapping samples to obtain normalized feature vectors;

[0079] A multi-parameter nonlinear regression equation system is established based on normalized eigenvectors, and the multi-parameter nonlinear regression equation system is embedded into a particle swarm-neural network hybrid solver to obtain the probe station inverse kinematics model. The multi-parameter nonlinear regression equation system contains three sets of independent equations, each set of independent equations corresponding to the control parameters of the three degrees of freedom of the probe station. The particle swarm-neural network hybrid solver is used to alternately iteratively optimize the coefficients of the multi-parameter nonlinear regression equation system and the neural network weights.

[0080] Based on the position difference between adjacent test points on the global optimal path input by the inverse kinematics model of the probe station, the probe station displacement control quantity sequence is obtained. Each displacement control quantity in the probe station displacement control quantity sequence includes single-step angle change, displacement step size and rotation step size.

[0081] The probe trajectory control sequence is obtained by curve fitting and trajectory smoothing based on the probe stage displacement control sequence.

[0082] In this example, the spatial coordinate information of all test points is extracted sequentially from the globally optimal path. Each test point contains three-dimensional coordinate components, corresponding to the lateral position, longitudinal position, and out-of-plane height position within the wafer plane. This three-dimensional coordinate sequence describes the precise spatial target that the probe needs to reach point by point. The three-dimensional coordinate sequence is then used to construct a probe station motion mapping sample with the corresponding probe station three-axis control parameters. These parameters include the horizontal movement angle, vertical displacement, and rotation angle around its own axis. These three parameters collectively define the probe's attitude and motion state in the three-dimensional coordinate system. By pairing each set of test point coordinates with its corresponding control parameters, a probe station motion mapping sample set is constructed. The probe station motion mapping samples are preprocessed, including feature standardization and spatial transformation. Feature standardization aims to map numerical features of different dimensions to a unified scale range, eliminating the interference of dimensional differences on the model's learning ability, and enabling the model to more effectively capture the relative relationships between features. Spatial transformation, based on the actual mechanical coordinate system of the probe station, performs geometric adjustments such as rotation, translation, and axis alignment on the coordinates to ensure that the input features logically conform to the actual control logic of the device. After standardization and spatial alignment, the original coordinate and control parameter pairs are transformed into a set of normalized feature vectors. Each input vector corresponds to a probe operation target, containing coordinate information and the corresponding motion control output, exhibiting good learnability and numerical convergence characteristics. Based on these normalized feature vectors, a set of multi-parameter nonlinear regression equations is established to construct the inverse kinematics model of the probe station. This set of equations consists of three independent equations, corresponding to the horizontal angle control, vertical displacement control, and rotation angle control of the probe station, respectively. The goal of each equation is to output the control parameter value corresponding to a given three-dimensional coordinate input. Since the motion behavior of the probe station is inherently highly nonlinear, state-coupled, and dynamically constrained, this nonlinear regression problem is embedded in a particle swarm optimization-neural network hybrid solver for solution. This hybrid solver combines the global search capability of particle swarm optimization (PSO) with the nonlinear fitting capability of neural networks. Employing an alternating optimization strategy, it first uses PSO to perform an initial search for the parameters of the equation system in each iteration, then uses neural networks to adjust the gradient of the fitting residuals of each regression function, dynamically updating network weights and regression coefficients. This ensures the model accurately approximates the sampling points while possessing strong generalization ability. The constructed inverse kinematics model can automatically predict the corresponding probe control commands after inputting 3D coordinates, exhibiting real-time response and high robustness. After model establishment, it is applied to control strategy generation. The 3D position difference between each pair of adjacent test points is extracted from the globally optimal path; these differences reflect the physical movement required by the probe between two tests.The displacement difference is input into the inverse kinematics model, which outputs a corresponding sequence of control commands. Each control variable includes the magnitude of the horizontal angle change, the vertical displacement step size, and the rotation step size required for attitude adjustment. These control variables together constitute the action plan required for the probe to smoothly transition from one test point to the next target point, marking the first stage of converting discrete path points into an executable control flow. Post-processing of the displacement control sequence mainly includes curve fitting and trajectory smoothing. Curve fitting uses interpolation methods such as cubic splines or Bézier curves to connect discrete control points into a continuous curve, ensuring that the probe's movement path is continuously differentiable in geometric space, thereby avoiding sudden speed changes and excessive impact forces that could cause equipment accuracy loss or hardware wear. Trajectory smoothing constrains the acceleration and angle change rates of the curve, ensuring that the trajectory meets the kinematic and dynamic constraints of the probe platform at the physical execution level, such as maximum speed, maximum acceleration, and maximum rotation angle. Intermediate transition points are inserted when necessary for dynamic attitude correction, improving trajectory stability and motion safety. The final output probe trajectory control sequence has high spatial accuracy and balances the response characteristics of the mechanical structure and path continuity during temporal execution.

[0083] Reference Figure 2 This embodiment provides a wafer probe station detection path planning device, including:

[0084] Boundary enhancement module 1 is used to perform boundary enhancement processing on the wafer surface image to obtain the boundary enhancement input tensor;

[0085] Feature extraction module 2 is used to perform dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map containing the chip boundary and test point locations;

[0086] Attention filtering module 3 is used to perform channel attention and spatial attention filtering on the first feature map to obtain the second feature map;

[0087] Distribution feature analysis module 4 is used to perform chip distribution feature analysis on the second feature map and obtain the clustering results of test points;

[0088] Path planning module 5 is used to perform probe movement path planning based on the test point clustering results to obtain the globally optimal path;

[0089] Calculation module 6 is used to establish the probe station inverse kinematics model based on the global optimal path, and to calculate the probe trajectory control sequence based on the probe station inverse kinematics model.

[0090] In this embodiment, the specific implementation of each unit in the above device embodiment is described in the above method embodiment, and will not be repeated here.

[0091] Reference Figure 3 This invention also provides a computer device, which can be a server, and its internal structure can be as follows: Figure 3 As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0092] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0093] An embodiment of the present invention also provides a probe station, which is used to implement the steps of the method described in any of the above claims.

[0094] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0095] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for planning the detection path of a wafer probe station, characterized in that, include: Boundary enhancement processing is performed on the wafer surface image to obtain the boundary enhancement input tensor; Dual-path feature extraction is performed based on the boundary enhancement input tensor to obtain a first feature map containing the chip boundary and test point locations; The first feature map is filtered using channel attention and spatial attention to obtain the second feature map; The second feature map is subjected to chip distribution feature analysis to obtain test point clustering results. Specifically, this includes: semantic segmentation of the second feature map to obtain a binarized test point distribution map; performing connectivity region analysis on the binarized test point distribution map to obtain basic features of the chip region and test points, the basic features including center coordinates, area, and number of test points; constructing a chip distribution covariance matrix based on the basic features, and applying a spectral clustering algorithm to the chip distribution covariance matrix to obtain K chip clusters; calculating and sorting the test point density for each chip cluster to obtain test priority region division information, where the test point density is defined as the number of test points per unit area; and performing fine-grained clustering of test points in high-density regions based on the test priority region division information to obtain test point clustering results. Based on the clustering results of the test points, probe movement path planning is performed to obtain the globally optimal path; A probe station inverse kinematics model is established based on the global optimal path, and the probe trajectory control sequence is calculated based on the probe station inverse kinematics model.

2. The wafer probe station detection path planning method according to claim 1, characterized in that, The boundary enhancement processing of the wafer surface image to obtain the boundary enhancement input tensor includes: The wafer surface image is input into three convolutional layers with convolutional kernels of different sizes for feature extraction, resulting in three sets of feature maps; The three sets of feature maps are weighted and fused to obtain a fused feature map; Edge detection is performed on the fused feature map using boundary extraction operators to obtain a boundary feature map. The boundary extraction operators include the horizontal Sobel operator, the vertical Sobel operator, and the Laplacian operator. Adaptive threshold segmentation is performed on the boundary feature map to obtain a boundary enhancement map, and the boundary enhancement map is channel-stitched with the wafer surface image to obtain a boundary enhancement input tensor.

3. The wafer probe station detection path planning method according to claim 1, characterized in that, The dual-path feature extraction based on the boundary enhancement input tensor yields a first feature map containing the chip boundary and test point locations, including: The boundary enhancement input tensor is input into the encoder of the main path for multi-level feature extraction, resulting in four encoded feature maps at different scales. The encoder contains four downsampling modules. The decoder of the main path decodes four encoded feature maps at different scales to obtain the main path decoded feature map. The decoder contains four upsampling modules. Three feature maps of different levels are selected from the main path decoding feature map and input into the supervision path to obtain three supervision branch feature maps; Channel dimensionality reduction and feature extraction are performed on the three supervised branch feature maps respectively to obtain supervised feature maps. The supervised feature maps are then sizing and fused with the main path decoding feature map to obtain a fused feature representation. Semantic segmentation processing is performed on the fused feature representation to obtain a first feature map containing chip boundaries and test point locations.

4. The wafer probe station detection path planning method according to claim 1, characterized in that, The step of performing channel attention and spatial attention filtering on the first feature map to obtain the second feature map includes: Perform global average pooling and global max pooling operations on the first feature map to obtain channel-dimensional statistical features; The channel dimension statistical features are input into two fully connected layers for dimensionality reduction and dimensionality increase respectively, to obtain the channel attention weight vector; The first feature map is subjected to average pooling and max pooling along the channel dimension to obtain two single-channel feature maps. The two single-channel feature maps are concatenated and then input into a 7×7 convolutional layer to obtain a spatial attention weight map. The channel attention weight vector is multiplied by the first feature map by channel, and then multiplied element-wise by the spatial attention weight map to obtain the preliminary filtered feature map; Based on the position of the preliminary selected feature map in the network, residual attention mechanisms with different coefficients are applied to the shallow feature map and the deep feature map respectively to obtain the second feature map, wherein the residual attention mechanism is used to preserve chip layout details and test point distribution information.

5. The wafer probe station detection path planning method according to claim 1, characterized in that, The step of performing probe movement path planning based on the clustering results of the test points to obtain the globally optimal path includes: Based on the clustering results of the test points, the density distribution features and spatial distribution features of the test points in each cluster are extracted to obtain the feature vector matrix; A nonlinear weighted calculation is performed on the feature vector matrix to obtain the path planning objective function; The clustering results of the test points are divided into multiple density level regions, and a test priority weight is assigned to each density level region to obtain a hierarchical detection sequence; Based on the layered detection sequence, a local sub-path is constructed for each density level region's test points to obtain the optimal path set within the region; Multi-objective particle swarm optimization is performed on the optimal path set within the region to obtain inter-regional path connection schemes; Based on the inter-regional path connection scheme, adaptive annealing-taboo search hybrid optimization is performed on the global path to obtain the globally optimal path.

6. The wafer probe station detection path planning method according to claim 1, characterized in that, The step of establishing an inverse kinematics model of the probe station based on the globally optimal path, and calculating the probe trajectory control sequence based on the inverse kinematics model of the probe station, includes: The test point three-dimensional coordinate sequence is extracted from the global optimal path, and the three-dimensional coordinate sequence is combined with the corresponding probe station three-axis control parameters to construct a probe station motion mapping sample. The probe station three-axis control parameters include the probe station's horizontal movement angle, vertical displacement, and rotation angle. The probe station motion mapping samples are subjected to feature standardization and spatial transformation to obtain normalized feature vectors; A multi-parameter nonlinear regression equation system is established based on the normalized feature vector, and the multi-parameter nonlinear regression equation system is embedded into a particle swarm-neural network hybrid solver to obtain the probe station inverse kinematics model. The multi-parameter nonlinear regression equation system contains three sets of independent equations, each set of independent equations corresponding to the control parameters of the three degrees of freedom of the probe station. The particle swarm-neural network hybrid solver is used to alternately iteratively optimize the coefficients of the multi-parameter nonlinear regression equation system and the neural network weights. Based on the position difference between adjacent test points on the global optimal path input by the inverse kinematics model of the probe station, a sequence of probe station displacement control quantities is obtained. Each displacement control quantity in the sequence of probe station displacement control quantities includes a single-step angle change, a displacement step size, and a rotation step size. Based on the probe stage displacement control sequence, curve fitting and trajectory smoothing are performed to obtain the probe trajectory control sequence.

7. A wafer probe station detection path planning device, characterized in that, For implementing the wafer prober detection path planning method according to any one of claims 1 to 6, the wafer prober detection path planning device comprises: The boundary enhancement module is used to perform boundary enhancement processing on the wafer surface image to obtain the boundary enhancement input tensor; The feature extraction module is used to perform dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map containing the chip boundary and the test point location; An attention filtering module is used to perform channel attention and spatial attention filtering on the first feature map to obtain a second feature map; The distribution feature analysis module is used to perform chip distribution feature analysis on the second feature map to obtain the test point clustering results; The path planning module is used to perform probe movement path planning based on the clustering results of the test points to obtain the globally optimal path; The calculation module is used to establish a probe station inverse kinematics model based on the global optimal path, and to calculate the probe trajectory control sequence according to the probe station inverse kinematics model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the wafer probe station detection path planning method according to any one of claims 1 to 6.

9. A probe station, characterized in that, The probe station is used to implement the steps of the wafer probe station detection path planning method as described in any one of claims 1 to 6.

Images