Wafer probe station detection path planning method, wafer probe station detection path planning device, wafer probe station detection path planning equipment and probe station

Abstract

The invention relates to the technical field of probe detection, and discloses a wafer probe station detection path planning method, device and equipment and a probe station. The method comprises the following steps: performing boundary enhancement processing on a wafer surface image to obtain a boundary enhancement input tensor; executing dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map including a chip boundary and a test point position; performing channel attention and space attention screening on the first feature map to obtain a second feature map; performing chip distribution feature analysis on the second feature map to obtain a test point clustering result; probe moving path planning is executed, and a global optimal path is obtained; and establishing a probe station inverse kinematics model based on the global optimal path, and calculating a probe track control sequence according to the probe station inverse kinematics model. According to the method, the wafer surface chip boundary feature extraction capability is enhanced, the stability and smoothness of the probe in the high-speed test process are ensured, and the probe positioning precision is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of probe detection, and particularly to a method, device, equipment and probe station for detecting path planning of a wafer probe station. Background Art

[0002] Traditional wafer probe station detection methods mainly use the serpentine scanning method and the spiral scanning method to plan the probe movement path. These methods do not fully consider the actual distribution characteristics of the chips on the wafer surface, often resulting in too long probe movement distance and low test efficiency. At the same time, with the increase of wafer size and the complexity of chip layout, the problem of insufficient chip boundary recognition accuracy becomes increasingly prominent, resulting in probe positioning deviation and affecting test accuracy.

[0003] The prior art pays insufficient attention to the path planning of high-density chip areas and cannot implement differential test strategies for different density areas, resulting in unreasonable test time allocation and wasting precious test resources. In addition, traditional path planning algorithms are difficult to adapt to the high non-linearity and uncertainty of the chip layout on the wafer surface, lack in-depth analysis of the chip test point distribution, and fail to perform intelligent path optimization according to the chip layout characteristics. Summary of the Invention

[0004] The main purpose of the present invention is to provide a method, device, equipment and probe station for detecting path planning of a wafer probe station. The present 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 greatly improves the probe positioning accuracy.

[0005] To achieve the above purpose, the present invention provides a method for detecting path planning of a wafer probe station, including the following steps: Perform boundary enhancement processing on the wafer surface image to obtain a boundary enhancement input tensor; Perform dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map including chip boundaries and test point positions; Perform channel attention and spatial attention screening on the first feature map to obtain a second feature map; Perform chip distribution feature analysis on the second feature map to obtain a test point clustering result; Perform probe movement path planning based on the test point clustering result to obtain a globally optimal path; Establish an inverse kinematics model of the probe station based on the globally optimal path, and calculate a probe trajectory control sequence according to the inverse kinematics model of the probe station.

[0006] The present invention also provides a device for detecting path planning of a wafer probe station, including: A boundary enhancement module for performing boundary enhancement processing on the wafer surface image to obtain a boundary enhancement input tensor; A feature extraction module for performing dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map including the chip boundary and the test point positions; An attention screening module for performing channel attention and spatial attention screening on the first feature map to obtain a second feature map; A distribution feature analysis module for performing chip distribution feature analysis on the second feature map to obtain a test point clustering result; A path planning module for performing probe movement path planning based on the test point clustering result to obtain a globally optimal path; A calculation module for establishing an inverse kinematic model of the probe station based on the globally optimal path and calculating a probe trajectory control sequence according to the inverse kinematic model of the probe station.

[0007] The present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the method described in any one of the above are implemented.

[0008] The present invention also provides a probe station for implementing the steps of the method described in any one of the above.

[0009] In summary, by introducing dual-path supervision and attention screening technologies, the present invention significantly improves the chip boundary recognition accuracy, enables the wafer probe station to more accurately locate the test points, and effectively solves the problem of inaccurate chip boundary recognition in the prior art. By combining a multi-scale convolutional neural network and a boundary extraction operator, the ability to extract chip boundary features on the wafer surface is enhanced, the probe positioning is more accurate, and the test error rate is reduced. Based on the feature analysis method of the chip distribution covariance matrix and the spectral clustering algorithm, an in-depth understanding of the chip distribution characteristics on the wafer surface is achieved, and reasonable test area division can be performed according to the chip density characteristics. Through the density level area division and priority weight allocation strategy, the present invention implements a differentiated test scheme for different density regions, optimizes resource allocation, and greatly improves the test efficiency. By proposing a multi-objective particle swarm optimization and adaptive annealing-tabu search hybrid algorithm, the global optimization of the probe path on the wafer surface with a complex layout is realized, and the probe movement distance and test time are significantly reduced. By establishing an accurate inverse kinematic model of the probe station and combining the variable step size B-spline curve interpolation method for trajectory planning, the stability and smoothness of the probe during high-speed testing are ensured, and the probe positioning accuracy is greatly improved. Description of the Drawings

[0010] Figure 1It is a schematic diagram of the steps of a wafer probe station detection path planning method in an embodiment of the present invention; Figure 2 It is a structural block diagram of a wafer probe station detection path planning device in an embodiment of the present invention; Figure 3 It is a structural schematic block diagram of a computer device in an embodiment of the present invention.

[0011] The realization, functional characteristics and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. Specific embodiments

[0012] In order to make the object, technical solution and advantages of the present invention clearer, the present 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 only used to explain the present invention and are not used to limit the present invention.

[0013] Referring to Figure 1 , this embodiment provides a wafer probe station detection path planning method, including the following steps: S1, perform boundary enhancement processing on the wafer surface image to obtain a boundary enhancement input tensor; Among them, the wafer surface images are respectively input into convolutional layers with three different sizes of convolutional kernels for feature extraction. These three groups of convolutional layers respectively use standard convolutional kernels with sizes of 3×3, 5×5, and 7×7 for parallel feature extraction operations. On the basis of maintaining the original structure of the image, the spatial variation characteristics exhibited by the edges of the wafer chips at different scales are captured. Each convolutional layer is set with a unified stride parameter of 1, and the number of output channels is fixed at 32 to ensure that the convolutional outputs at different scales have consistent channel dimensions, facilitating subsequent feature fusion processing. Through the feature fusion module, weighted fusion operations are performed on the three groups of feature maps. This module implements fusion by bitwise weighted summation, where the weight coefficients corresponding to the outputs of the three convolutional kernels are respectively initialized as 0.4, 0.35, and 0.25. During the fusion process, the weight coefficients are dynamically adjusted through the backpropagation mechanism of the neural network, enabling the model to adaptively select the most representative scale feature information. After fusion, a fusion feature map integrating multi-scale edge perception ability is obtained. On the basis of the fusion feature map, a boundary extraction operator is introduced to perform edge detection operations, including a group of Sobel operators for extracting horizontal edges, a group of Sobel operators for extracting vertical edges, and a group of Laplacian operators for extracting omnidirectional edges. The results output by these three edge operators are weighted and fused according to a weight ratio of 1:1:0.8 to generate a boundary feature map. The adaptive threshold segmentation algorithm is applied to the boundary feature map to dynamically calculate the threshold T suitable for the current image feature distribution. Its calculation formula is T = μ + k×σ, where μ is the global mean of the boundary feature map, σ is the standard deviation, and k is the adjustment coefficient, initially set to 1.5. Binaryzation operation is performed on the boundary feature map through this threshold, effectively separating the edge significant regions from the background to generate a boundary enhancement map. The boundary enhancement map is stitched with the original wafer surface image along the channel dimension to form a multi-channel input tensor, constituting a boundary enhancement input tensor for subsequent deep feature extraction and path planning analysis.

[0014] S2, 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 positions; Specifically, the boundary enhancement input tensor is input into the encoder part of the main path structure. The encoder is based on the improved U-Net architecture and contains four downsampling modules. Each module consists of two 3×3 convolutional layers, a batch normalization layer and a maximum pooling layer, which can effectively compress the spatial size and extract semantic information at different levels. Through layer-by-layer downsampling processing, the multi-level feature expression from local edge texture to global structural contour is gradually captured, and the encoded feature maps at four different spatial scales are output. These feature maps correspond to the manifestation of wafer images under different receptive fields, and have strong hierarchical and spatial distribution recognition capabilities. The above four encoded feature maps are decoded layer by layer by the main path decoder. The decoder also contains four upsampling modules, each of which includes a transposed convolutional layer and two 3×3 standard convolutional layers, and is equipped 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 jump-connected with the encoded feature map of the corresponding scale downsampling stage to enhance the edge and structure restoration capabilities and reconstruct a high-resolution main path decoding feature map. While the main path completes the decoding process, three representative feature maps of different levels are selected from the four feature maps output by the encoder stage of the main path as the input source of the supervision path, and they are sent to three independent branches of the supervision path respectively. Each supervision branch contains a 1×1 convolution layer for channel dimension reduction, a 3×3 convolution layer for local feature extraction, and an upsampling layer for size alignment to ensure that feature maps from different depths can be uniformly mapped to the main path decoding feature map. Figure 1 The dimensionality reduction operation effectively reduces the computational complexity, while allowing the shallow detail features and deep semantic features to be extracted and enhanced in a lightweight manner. The three supervisory branch feature maps are fused with the decoded feature map of the main path through a resizing operation. The fusion process is completed by feature concatenation or bitwise weighted summation to construct a fused feature representation. The fused feature representation is input into the semantic segmentation module for processing. The semantic segmentation module uses a fully convolutional structure for pixel-by-pixel classification. According to the structural boundary clues and test point response information in the fused features, a high-resolution feature map containing the wafer chip boundary and the test point position is output, which is the first feature map.

[0015] S3, performing channel attention and spatial attention screening on the first feature map to obtain a second feature map; It should be noted that for the significance modeling operation of the first feature map in the channel dimension, global average pooling and global max pooling are respectively performed on it. These two operations compress along the spatial dimension and retain the global statistical features of each channel. Among them, average pooling reflects the overall response intensity of the channel, while max pooling emphasizes the response peak of the most significant region. The two jointly constitute a statistical vector describing the importance of the channel. The two statistical vectors are respectively input into a network structure composed of two fully connected layers. The dimensionality reduction and compression of 1 / 16 of the number of channels are performed through FC1. After activation by ReLU, it is then input into FC2 to complete the dimensionality increase operation, and the vector is restored to the original number of channels to form a channel attention weight vector with fine expression. After the output results of the two channel paths are added on a per-channel dimension basis, they are normalized through the Sigmoid function to obtain a channel weight coefficient ranging from 0 to 1, which is used to characterize the importance of each channel. Average pooling and max pooling are respectively performed on the first feature map along the channel dimension to generate two single-channel spatial feature maps. The former focuses on the average response region of all channels, and the latter emphasizes the spatial distribution information dominated by the maximum value in the channel. The two respectively capture the spatial significant region features from the perspectives of balance and extreme values. These two single-channel images are concatenated along the channel direction into a two-dimensional structure diagram and input into a convolutional kernel with a size of 7×7 for convolution operation to capture the local spatial relationship under a larger receptive field. The convolution result is activated through the Sigmoid function to generate a spatial attention weight map with the same spatial dimension as the original image, identifying the regions that need to be retained and suppressed at the spatial level. The channel attention weight vector is multiplied by the first feature map channel by channel along the channel dimension to adjust the response intensity of each channel and retain the feature channels with rich information and relevant to the task. Subsequently, the result is multiplied by the spatial attention weight map pixel by pixel according to the element position to enhance the response ability of the significant region at the spatial level and weaken the interference of the useless region. Through this combined attention mechanism, dual significance screening in the channel dimension and the spatial dimension is completed, and a preliminary screening feature map after channel screening and spatial screening is output. According to the position of the preliminary screening feature map in the network, residual attention mechanisms with different coefficients are respectively applied to the shallow feature map and the deep feature map, and the screening intensity is dynamically adjusted according to the position of the feature map in the network. For the shallow feature map close to the input end of the network, since it focuses on capturing local edges and chip contour information, the original structural details are retained, and conservative feature enhancement is achieved through a lower-intensity residual enhancement strategy; while for the deep feature map close to the output end, it pays more attention to the expression of high-semantic targets such as test points, and a stronger residual enhancement strategy is adopted to significantly improve the expression effect of high-level abstract features. By applying differential enhancement to the shallow and deep feature maps respectively, the gradual strengthening and focusing of the information flow from low-level details to high-level semantics in the entire network structure are realized, and the second feature map is obtained.

[0016] S4. Analyze the chip distribution characteristics of the second feature map to obtain the test point clustering result; Specifically, input the second feature map into a semantic segmentation network. This network has the ability to identify regions at the pixel level, can accurately divide various semantic regions in the wafer image, and output a binary distribution map that only contains the test point region. In this image, the pixel values of all identified test points are 1, and the rest are 0. This operation calibrates the spatial distribution of the test points on the wafer surface. After completing the semantic segmentation, perform a connected component analysis on the test point distribution map to identify each connected test point cluster, and extract representative basic feature information for each cluster, including the central position of the region, the spatial occupied area, and the number of test points within the region. The central coordinates reflect the geometric position of the chip or region on the entire wafer, the area is used to describe its spatial coverage, and the number of test points is an important parameter to measure the detection complexity of this region. After integrating these data, obtain the basic descriptions of multiple independent chip regions. On this basis, analyze the spatial distribution structure of the chips 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 to obtain the chip distribution covariance matrix. Apply the spectral clustering algorithm to the chip distribution covariance matrix, where the number of clusters is adaptively determined according to the relative closeness between regions, thereby dividing the entire wafer surface into several region groups with similar structures and functions. Statistically analyze the test point density within each cluster. The test point density refers to the number of test points contained in a unit area of a certain chip cluster. The higher the density of a region, the greater its detection complexity and the more concentrated the test time, so it is regarded as a high-priority test region. By sorting the test densities of all clusters, establish a test priority division standard, so that the probe preferentially processes the most complex and dense test regions during actual movement, improving the overall test efficiency. For high-priority regions, perform a more refined test point clustering process. Select a clustering algorithm that is sensitive to local dense structures to refine the test points within each high-density region, forming multiple smaller-grained and more compact test point groups, and finally obtain a test point clustering result with clear boundaries, reasonable structure, and test priorities.

[0017] S5. Based on the test point clustering result, perform probe movement path planning to obtain the globally optimal path; Among them, based on the clustering results of the test points, the density distribution characteristics and spatial distribution characteristics of each clustering region are extracted. The density distribution characteristics are used to reflect the density of the test points in each region, and the spatial distribution characteristics describe the arrangement pattern and geometric expansion characteristics of the test points within the plane range. These information are uniformly encoded into several description vectors to form a multi-dimensional feature vector matrix. A non-linear weighted calculation is performed on the feature vector matrix to establish a path planning objective function that comprehensively considers the influence of multiple factors. Among them, multiple indicators such as moving distance, test order, region complexity, and path coherence are integrated to form an optimization evaluation criterion that reflects the balance relationship between the probe movement efficiency and the test coverage efficiency. The test point clustering results are divided into multiple density-level regions, and different test weights are configured according to the detection priorities represented by each level. The weights determine the priority processing order and access frequency of the high-density regions in the probe path planning, thereby constructing a hierarchical detection sequence system. Based on this hierarchical structure, path construction operations are separately performed for the test points within each density-level region. Using heuristic or fast iterative shortest path search strategies, corresponding local sub-paths are generated for different regions, and the optimal path set within each region is selected according to criteria such as path compactness, rationality of point-to-point connection, and local test cost. This set represents the optimal test path structure within a fixed region range. After completing the local path construction, the scattered optimal sub-paths are combined into an overall path system, and a multi-objective particle swarm optimization algorithm is introduced for global connection planning. This optimization process aims to minimize the connection cost between regions, the overall path continuity, and the path overlap degree. It continuously iterates on the path connection scheme between regions, and searches for the optimal connection strategy by dynamically adjusting the particle positions and the state evolution process to form a preliminary global path structure that serially connects the local optimal paths in an orderly manner. A hybrid optimization strategy that combines an adaptive annealing algorithm and a tabu search mechanism is performed on the preliminary connection path. By introducing a probability jump-out mechanism through the simulated annealing method to avoid falling into local minima, and at the same time combining the historical state record and path replacement strategy of the tabu search, the path structure is effectively inhibited from falling into repeated updates and cyclic redundancy, thereby completing the iterative optimization of the probe movement path and the efficient fine-tuning of local details on a global scale. The globally optimal path is obtained.

[0018] S6. Based on the globally optimal path, an inverse kinematic model of the probe station is established, and the probe trajectory control sequence is calculated according to the inverse kinematic model of the probe station.

[0019] Specifically, a three-dimensional coordinate sequence of test points is extracted from the globally optimal path, and this sequence describes the spatial positions that the probe visits sequentially on the wafer surface according to 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 real control responses, where the control parameters include the moving angle of the probe station in the horizontal direction, the precise displacement in the vertical direction, and the rotation angle to ensure attitude consistency. Feature normalization and spatial transformation are performed on the motion mapping samples of the probe station. By unifying the scale range, the coordinate data and control parameters in each sample are transformed into a normalized numerical interval, and rotation alignment and coordinate system mapping are performed on the premise of keeping the geometric structure unchanged, so as to construct a set of normalized feature vectors with a unified reference framework and numerical computability. Based on the normalized feature vectors, a multi-parameter nonlinear regression equation set describing the control behavior of the probe station is constructed. This equation set contains three groups of non-linear functions that are independent of each other but mathematically related. Each group of equations corresponds to the control instructions for the three degrees of freedom of the probe station, that is, the calculation expressions of the horizontal angle, vertical displacement, and rotation angle. To improve the solution accuracy and model convergence performance, the regression equation set is embedded in a hybrid solver jointly constructed by the particle swarm optimization algorithm and the neural network. Through the global search ability of the particle swarm in the continuous space and the local accuracy advantage of the neural network in function fitting, the coefficients of the regression equation set and the weight parameters of the neural network are alternately iteratively updated, so that the entire inverse kinematics model not only has the ability to accurately model complex motion relationships but also has good convergence and adaptability. After the model is constructed, the spatial position difference between adjacent test points on the globally optimal path is used as the input and is sequentially fed into the inverse kinematics model to calculate the corresponding probe station displacement control quantity sequence. Each control quantity describes the single-step horizontal angle change, longitudinal displacement step, and rotation adjustment step required when moving from the current position to the next test point. These control quantities are combined item by item to form a continuous control instruction sequence, which constitutes the control framework for the entire process of the probe from the starting point to the ending point. To improve the smoothness of the probe operation and the continuity of the mechanical response, a curve fitting and trajectory smoothing processing mechanism is introduced based on the displacement control quantity sequence. Through cubic spline interpolation or high-order Bezier fitting algorithms, the curvature of the connection segments between the control quantities is optimized and redundancy is suppressed, so that the speed, acceleration, and angle changes of the probe during execution are smoother, effectively avoiding system oscillation, error accumulation, and mechanical fatigue. The probe trajectory control sequence is output through the above process.

[0020] In one example, boundary enhancement processing is performed on the wafer surface image to obtain a boundary enhancement input tensor, including: The wafer surface image is respectively input into convolutional layers with three different sizes of convolutional kernels for feature extraction to obtain three groups of feature maps; Weighted fusion processing is performed on the three groups of feature maps to obtain a fused feature map; Edge detection is performed on the fused feature map using edge extraction operators to obtain a boundary feature map. The edge extraction operators include a horizontal Sobel operator, a vertical Sobel operator, and a 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-concatenated with the wafer surface image to obtain a boundary enhancement input tensor.

[0021] In this example, the original wafer surface image obtained by acquisition is input into three convolutional layers with different receptive fields to obtain three sets of feature maps. These three convolutional layers use convolutional kernels with sizes of 3×3, 5×5, and 7×7 respectively to process the input image, capturing the structural information and boundary details in the wafer image from different scales. The smaller-sized convolutional kernels can sensitively perceive local texture changes, while the larger-sized convolutional kernels help capture broader background contours and shape structures. The three sets of feature maps are processed by weighted fusion. A weighted fusion mechanism is adopted, and different fusion weights are set for each set of feature maps to dynamically adjust the proportion of different-scale features in the overall expression. The weights are preset with initial values and then continuously optimized through the backpropagation strategy during network training, enabling the model to autonomously learn which type of features contribute more to boundary recognition and achieving optimal feature fusion. The fusion method uses pixel-by-pixel weighting, where the feature values at corresponding positions in the three channels are weighted and summed according to the set weights, thereby retaining the response advantages of multi-scale features in the fusion map while eliminating redundant information and inefficient activations, generating a fused feature map. An edge extraction operator is used to perform edge detection on the fused feature map. The edge extraction operator includes the horizontal Sobel operator, the vertical Sobel operator, and the Laplacian operator. The Sobel operator is an operator that detects edges by calculating the gray-scale gradient of the image. The horizontal Sobel operator is used to detect horizontal boundary changes in the image, while the vertical Sobel operator is used to extract vertical edge contours. The two work together to outline the main boundary direction change trend. The Laplacian operator is a second-order derivative operator with omnidirectional detection capabilities, responding to all edge changes in the image and being good at capturing structural mutation points and high-frequency boundary signals. In practical applications, the fused feature map is separately input into the above three edge detection operators for independent calculation to obtain three edge response maps in corresponding directions. Subsequently, these three maps are weighted and fused, and the fusion weights are set or trained and optimized according to the strength distribution of edge responses in the actual scenario to ensure that the final boundary feature map can retain the true chip boundary to the greatest extent and suppress noise and pseudo-edges, obtaining a boundary feature map. The boundary feature map is binarized, converting the edge map with continuous pixel values into a binary image that only contains edges and non-edges. An adaptive threshold segmentation method is adopted, dynamically generating an optimal segmentation threshold according to the statistical characteristics of the image itself to adapt to the variation characteristics of different wafer images in terms of brightness, contrast, and edge strength, avoiding over-segmentation or under-segmentation problems caused by using a fixed threshold. During the calculation process, the pixel average value of the entire boundary map is selected as the basis, and the final threshold is dynamically determined in combination with the pixel standard deviation and adjustment parameters. The pixel values in the boundary feature map are discriminated using this threshold. Those higher than the threshold are judged as boundaries, and those lower than the threshold are judged as non-boundaries, forming a boundary-enhanced map with a stable boundary contour structure.Perform channel - level splicing on the boundary - enhanced image and the original wafer image, and input tensors with different information dimensions into the neural network in a parallel structure to form a multi - channel boundary - enhanced input tensor.

[0022] In one example, perform dual - path feature extraction based on the boundary - enhanced input tensor to obtain a first feature map containing the chip boundary and the test - point positions, including: Input the boundary - enhanced input tensor into the encoder of the main path for multi - level feature extraction to obtain four encoded feature maps with different scales. The encoder contains four down - sampling modules; Decode the four encoded feature maps with different scales through the decoder of the main path to obtain the main - path decoded feature map. The decoder contains four up - sampling modules; Select three feature maps at different levels from the main - path encoded feature maps and input them into the supervision path to obtain three supervised - branch feature maps; Perform channel dimension reduction and feature extraction on the three supervised - branch feature maps respectively to obtain the supervised feature map, and after adjusting the size of the supervised feature map, fuse it with the main - path decoded feature map to obtain the fused feature representation; Perform semantic segmentation processing on the fused feature representation to obtain a first feature map containing the chip boundary and the test - point positions.

[0023] In this example, the boundary-enhanced input tensor is fed into the encoder structure of the main path. The encoder 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 convolutional kernels to extract local spatial features. The batch normalization layer is used to accelerate convergence and stabilize the training process. The max pooling operation scales down the spatial dimensions proportionally, thereby gradually reducing the resolution and increasing the receptive field of the features. During the four downsampling processes, the spatial size of the feature maps decreases successively, while the channel dimension gradually increases. The encoded feature maps extracted layer by layer retain the structural semantic information at different scales. The shallower feature maps pay more attention to edge textures and spatial localization, while the deeper ones form an abstract representation of the overall distribution of the chip arrangement pattern and test point structure. This layer-by-layer extraction mechanism enables the model to comprehensively understand the spatial composition of the wafer image from local to global. After encoding, the encoded feature maps at four scales are fed into the decoder structure in sequence. The decoder consists of four upsampling modules. Each module includes a transposed convolutional operation, two standard convolutional layers, and a batch normalization layer. The transposed convolutional operation is used to gradually restore the size of the feature maps to the original image scale. The convolutional operations are used to fuse the information of the current layer with the interpolated features during the upsampling process. In addition, through the skip connection mechanism, the features of the corresponding levels in the encoder are directly introduced into the decoder to enhance the restoration ability of edge information and the level of detail retention. After four rounds of upsampling, the decoded feature map of the main path is obtained. Three different levels of outputs with representativeness and feature distinctiveness are selected from the four feature maps generated in the main path encoding stage and introduced as inputs into three independent branches of the supervision path respectively. These three groups of feature maps represent shallow, middle, and deep features respectively, and the spatial distribution and local contour information they carry are complementary. The supervision path takes them as parallel processing objects and performs a series of operations in their respective branches to extract task-relevant feature representations. Each supervision branch reduces the channel dimension through a one-to-one convolutional operation. This operation can reduce redundant information and compress the computational complexity, so that features at different levels will not lose alignment due to inconsistent channels during subsequent processing. After dimensionality reduction, standard convolution processing is performed to enhance the feature discrimination ability, and finally, the supervised feature maps output at different levels are adjusted to the same spatial size as the decoded feature map of the main path by upsampling, so as to provide an aligned input data structure for subsequent fusion processing. After completing the size adjustment, these three supervised feature maps are fused with the decoded feature map of the main path. This fusion process adopts the method of channel concatenation or weighted summation, and a more suitable fusion strategy is selected according to the actual training situation. By fusing the fine-grained, multi-level boundary structures in the supervision path with the global, coarse-to-fine semantic information in the main path, the resulting fused feature representation has good boundary clarity in the spatial structure and has the ability to understand the macroscopic chip position and grasp the microscopic test point distribution at the semantic level.The fused feature representation is input into the semantic segmentation module. This module uses a fully convolutional structure for per-pixel classification and combines multi-level convolutional kernels with a context extraction mechanism to perform boundary judgment and target point recognition for each pixel point in the wafer image. During the training process, a joint optimization objective combining cross-entropy loss and region overlap loss is adopted, enabling the model to not only focus on the overall classification accuracy but also improve the boundary fitting accuracy in the extraction of small-region test points. The finally output first feature map is a structure diagram with the same spatial dimension as the original. Figure 1 It depicts the contour boundary of the wafer chip and at the same time marks the specific positions of all test points in the image.

[0024] In one example, channel attention and spatial attention screening are performed on the first feature map to obtain a second feature map, including: Global average pooling and global max pooling operations are performed on the first feature map to obtain channel-dimension statistical features; The channel-dimension statistical features are respectively input into two fully connected layers for dimensionality reduction and dimensionality increase processing to obtain channel attention weight vectors; Average pooling and max pooling are respectively performed on the first feature map 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 vectors are multiplied by the first feature map channel by channel, and then multiplied by the spatial attention weight map element by element to obtain a preliminary screening feature map; According to the position of the preliminary screening feature map in the network, residual attention mechanisms with different coefficients are respectively applied to the shallow feature map and the deep feature map to obtain the second feature map, where the residual attention mechanism is used to retain the chip layout details and test point distribution information.

[0025] In this example, the first feature map is compressed in the spatial dimension. Through global average pooling and global maximum pooling, the responses in each channel dimension are statistically calculated as two one-dimensional vectors. Global average pooling reflects the overall response intensity trend in the channel, while global maximum pooling captures the distribution of the strongest response points in the channel. These two sets of channel statistical features are respectively input into two fully connected networks with the same structure. The fully connected network consists of two linearly transformed layers in series. The first layer plays a role in dimensionality reduction, reducing the original number of channels to a fixed ratio. The compressed intermediate representation is processed by a non-linear activation function and then input into the second fully connected layer for dimensionality increase operation to restore to the original number of channels. Through the dimensionality reduction - dimensionality increase structure, the model focuses on the correlation and dependency structure between each channel during the learning process, forming a set of attention weight vectors with channel perception ability. The two channel attention weight vectors are element-wise added and fused, and an activation function is used to compress them into the normalized range, outputting a set of attention coefficients consistent with the original number of channels. At the same time, 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 an aggregation operation on it in the channel dimension. Two single-channel two-dimensional feature maps are generated through average pooling and maximum pooling in the channel dimension. These two maps respectively reflect the mean distribution of the cumulative responses across all channels and the position structure of the extreme responses, providing complementary perspectives on the chip layout and structural contour at the spatial level. Subsequently, these two maps are concatenated in the channel dimension into a two-channel intermediate tensor and input into a two-dimensional convolutional layer with a convolutional kernel size of 7×7 for local spatial relationship modeling. This convolutional operation has a large receptive field and can capture microstructural features such as boundary lines, regional contours, and spatial repetitiveness in the neighborhood. The convolutional output is non-linearly activated to obtain a spatial attention weight map with the same size as the input map, which is used to indicate the attention intensity at each pixel position within the two-dimensional spatial range. 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 channel-reweighted intermediate feature map is then multiplied by the spatial attention weight map pixel by pixel to form a preliminary screening feature map that combines channel dimension selectivity and spatial dimension attentiveness. To enhance the adaptability of the feature map to the task objective and suppress the information weakening that occurs during network propagation, a differential residual attention mechanism is applied according to the position of the current feature map in the entire network structure.For the feature maps in the shallow network, its main task is to retain the chip contours with clear details and distinct boundaries in the wafer image. Therefore, a residual enhancement strategy with a smaller ratio is adopted, and the original input feature map and the screened feature map are superimposed residually with a smaller coefficient to enhance the expression of significant regions while maintaining the original texture. For the feature maps in the deep network, which carry more global expressions at the semantic level such as the distribution of test points, density status, and layout structure, a residual superposition strategy with a larger ratio is adopted to amplify the local anomalies, aggregated high-density regions, and spatial center features shown in the screened feature maps, thereby improving the model's identification ability for the test area. The shallow and deep feature maps generate complete second feature maps after being corrected by the corresponding residual coefficients.

[0026] In one example, the chip distribution feature of the second feature map is analyzed to obtain the test point clustering result, including: Semantic segmentation is performed on the second feature map to obtain a binary test point distribution map; Connected component analysis is performed on the binary test point distribution map to obtain the basic features of the chip area and test points. The basic features include the center coordinates, area, and the number of test points; A chip distribution covariance matrix is constructed based on the basic features, and the spectral clustering algorithm is applied to the chip distribution covariance matrix to obtain K chip clusters; The test point density is calculated and sorted for each chip cluster to obtain the test priority area division information, where the test point density is defined as the number of test points per unit area; Based on the test priority area division information, fine-grained clustering is performed on the test points in the high-density area to obtain the test point clustering result.

[0027] In this example, the second feature map is input into the semantic segmentation module to identify the regions of test significance in the wafer image and accurately mark the positions of the test points therein. The semantic segmentation network adopts a fully convolutional architecture and is combined with a multi-scale feature fusion strategy to obtain high semantic information while maintaining the spatial resolution. Through the pixel-level classification ability of the network, a classification image with the same size as the input image is output. The output image is converted into a binary test point distribution map, in which all pixels identified as test points are uniformly marked as 1, and the remaining regions are marked as 0, forming a structure diagram expressing the spatial distribution of test points. Connected component analysis is performed on this binary image. This operation is based on the connectivity principle that adjacent pixel values in the image are the same, and all independent regions composed of test point pixels are identified. Each connected component corresponds to a physical chip unit or a sub-test area. By traversing each connected component and extracting its basic geometric features, basic data descriptions about the chip distribution and test point clusters are obtained. Among them, the center coordinates are calculated through the pixel centroid to indicate the position of the region on the entire wafer plane; the area is determined by calculating the total number of pixels in the 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 in the region, revealing the density and complexity of the internal test tasks in this region. These features together constitute the basic data set for each chip region. After standardizing the basic feature vectors of all chip regions, a covariance matrix is constructed. This matrix is used to measure the comprehensive similarity between any two regions in terms of spatial position, area scale, and test density. On this basis, a spectral clustering algorithm is introduced to process this covariance matrix, mapping the data points to a low-dimensional space, and using the graph structure between the data for clustering in this space. By performing eigen-decomposition on the matrix and constructing a graph Laplacian matrix, the model identifies regions with similar structures or test intensities distributed on the wafer and groups them into the same category, thereby dividing the original chip distribution structure into K clusters. Each cluster represents a set of physically or logically closely related chip regions. The value of K is adaptively adjusted through unsupervised metrics such as the silhouette coefficient to balance the number of clusters and the quality of clustering. After obtaining these clustering results, to determine the priority in path planning, the test point density of each chip cluster is calculated. The density is obtained through the ratio between the total number of test points within the cluster and the total area it covers. The greater the density, the more concentrated the test points and the more frequent the path switching in this region, so it is regarded as a high-complexity and high-priority test region. After sorting all the clusters in descending order of test density, a set of test priority region division information is generated. This information is used to determine the order of regions that the probe enters first in the path optimization and task scheduling phases, thereby improving the overall test efficiency and reducing unnecessary movements.For areas with extremely high test density, the internal test points are distributed very densely and without obvious structural patterns. Therefore, finer-grained clustering processing is performed on the high-density areas. Introduce a fine-grained clustering method, such as density-based regional clustering technology, to identify subsets of test points that are closely concentrated in the local space, and naturally divide the sparse points or outliers on the edge into the periphery or eliminate them. By performing fine-grained division operations on the test points in the high-density area, the complex structure is transformed into several manageable sub-blocks, and local optimization space units are provided for subsequent path planning, enabling the probe to form short-range and efficient sub-paths in each sub-block during the test task, and then integrating the sub-paths into the overall test path.

[0028] In one example, after the test point clustering result, probe movement path planning is performed to obtain the global optimal path, including: Extract the test point density distribution characteristics and spatial distribution characteristics of each cluster based on the test point clustering result to obtain the feature vector matrix; Perform non-linear weighted calculation on the feature vector matrix to obtain the path planning objective function; Divide the test point clustering result into multiple density-level areas, and assign test priority weights to each density-level area to obtain a hierarchical detection sequence; Based on the hierarchical detection sequence, construct local sub-paths for the test points in each density-level area to obtain the set of optimal paths within the area; Perform multi-objective particle swarm optimization on the set of optimal paths within the area to obtain the path connection scheme between areas; Based on the path connection scheme between areas, perform adaptive annealing-tabu search hybrid optimization on the global path to obtain the global optimal path.

[0029] In this example, structural feature extraction is performed on each clustering region to construct a feature vector matrix with numerical description capabilities. For each test point clustering, calculate the average distribution density of the internal test points per unit area, and extract its spatial layout features, including centroid coordinates, geometric boundary shape, main direction expansion angle, and indicators such as the average distance, maximum distance, and minimum distance between test points. These features together form a complete multi-dimensional vector, reflecting the complexity within the clustering and embodying the relative position and density characteristics of the region in the overall wafer structure. Arrange the feature vectors of all clustering regions in rows to form a set of feature vector matrices. Perform weighted calculations on different feature dimensions in the feature vector matrix. Since the physical meanings of different feature dimensions are different and there are non-linear differences in their impacts on the path structure, the weighting method needs to be set in combination with actual test constraints. Adopt a non-linear mapping strategy to perform a function transformation on each eigenvalue according to the cost it generates in the actual path, and then add the transformed values according to the set weight coefficients to generate a comprehensive evaluation index for path planning. This index serves as the objective function in the path optimization process, covering basic requirements such as path length, turning complexity, and test point coverage efficiency, and comprehensively evaluating the density change, regional transition continuity, and control action switching frequency in the path, thereby guiding the optimization algorithm to converge towards the direction of short overall movement distance, reasonable test sequence, and high mechanical execution efficiency. After the objective function is constructed, re-structure the test point clustering results, and divide all clustering regions into multiple density-level regions. The division criterion is jointly judged based on the test point density and spatial scale of each clustering, and is divided into three categories: high-density area, medium-density area, and low-density area. Among them, the high-density area corresponds to the area where the test tasks are concentrated, and higher requirements are put forward for the control accuracy and time efficiency of the path, while the low-density area appropriately relaxes the local constraints of path planning. Based on the density-level division, different test priority weights are assigned to different density regions. The higher the weight, the higher the priority of the region in the test execution sequence, thus constructing a hierarchical detection sequence mechanism. Based on this hierarchical detection sequence, construct local paths for the test points within each density-level region. Within each region, use distance heuristic, nearest neighbor strategy, or graph search algorithm to combine all test points to generate multiple feasible test point access paths, and then select the local optimal path based on indicators such as path length, angle change, and point-to-point jump distance. This operation ensures the shortest path for point-to-point connection within the region, and at the same time considers the boundary smoothness and connectivity of the path when accessing the global path. Therefore, a set of connectable edge points needs to be reserved at the end of each local path for subsequent global path splicing. After all local paths are constructed, input the local paths as sub-task units into the multi-objective particle swarm optimization module. Particle swarm optimization not only considers the connection order between paths, but also considers objectives such as the turning cost, connection jump distance, overall path smoothness, and path execution time between each path segment.By globally evaluating the path sequence of each particle and continuously adjusting the particle decoding strategy and the information interaction mechanism between particles during the iteration process, it gradually converges to a set of path connection schemes that achieve an optimal balance between the global shortest distance and regional connection efficiency. This scheme is output in a specific regional connection order and connection point integration method to serve as the framework for the final path structure. Based on the determined inter-regional path connection scheme, a hybrid optimization mechanism combining adaptive annealing and tabu search is introduced to eliminate local discontinuous segments, redundant detour points, and repeated node access problems in the path. The adaptive annealing algorithm allows accepting inferior solutions to a certain extent to jump out of the local optimum, thereby increasing the global optimum convergence probability of the overall path structure; while tabu search avoids the search from falling into repeated loops and ineffective transformations by recording the historical solution state and prohibited regions. The combination of the two solves the problems of global optimization and local trap control simultaneously. During the optimization process, local segments of the path will undergo perturbation transformations, such as node swapping, path segment reversal, in-segment interpolation, etc. All transformed results are re-evaluated based on the objective function and selected whether to be retained according to the probability rule. The finally output path achieves overall optimality in multiple dimensions such as total path length, number of turns, sub-path connection stability, and execution coherence.

[0030] In an example, an inverse kinematic model of the probe station is established based on the globally optimal path, and a probe trajectory control sequence is calculated according to the inverse kinematic model of the probe station, including: Extracting the three-dimensional coordinate sequence of test points from the globally optimal path and constructing a probe station motion mapping sample with the three-dimensional coordinate sequence and the corresponding three-axis control parameters of the probe station. The three-axis control parameters of the probe station include the horizontal movement angle, vertical displacement amount, and rotation angle of the probe station; Performing feature standardization and space transformation on the probe station motion mapping sample to obtain a normalized feature vector; Establishing a multi-parameter nonlinear regression equation set based on the normalized feature vector and embedding the multi-parameter nonlinear regression equation set into a particle swarm-neural network hybrid solver to obtain an inverse kinematic model of the probe station. The multi-parameter nonlinear regression equation set contains three groups of independent equations, and each group of independent equations corresponds to the control parameters of three degrees of freedom of the probe station. The particle swarm-neural network hybrid solver is used to alternately iterate and optimize the coefficients of the multi-parameter nonlinear regression equation set and the neural network weights; Inputting the difference in the positions of adjacent test points on the globally optimal path based on the inverse kinematic model of the probe station to obtain a probe station displacement control quantity sequence. Each displacement control quantity in the probe station displacement control quantity sequence includes a single-step angle change amount, a displacement step length, and a rotation step length; Performing curve fitting and trajectory smoothing processing based on the probe station displacement control quantity sequence to obtain a probe trajectory control sequence.

[0031] In this example, the spatial coordinate information of all test points is sequentially extracted from the globally optimal path. Each test point contains three-dimensional coordinate components, corresponding to the lateral position, longitudinal position within the wafer plane, and out-of-plane height position respectively. This three-dimensional coordinate sequence describes the precise spatial targets that the probe needs to reach point by point. A motion mapping sample of the probe station is constructed by pairing the three-dimensional coordinate sequence with the corresponding three-axis control parameters of the probe station. The three-axis control parameters of the probe station include the horizontal movement angle, vertical displacement, and rotation angle around its own axis. These three parameters jointly define the posture and motion state of the probe in the three-dimensional coordinate system. By pairing each set of test point coordinates with their corresponding control parameters, a set of motion mapping samples of the probe station is constructed. Preprocessing is performed on the motion mapping samples of the probe station, including feature standardization and spatial transformation. The purpose of feature standardization is to map numerical features of different dimensions to a unified scale range, eliminate the interference of dimensional differences on the model's learning ability, and enable the model to more effectively capture the relative relationships between features. Spatial transformation performs geometric adjustments such as rotation, translation, and axis alignment on the coordinates according to the actual mechanical coordinate system of the probe station, ensuring that the input features conform to the real control logic of the device in terms of logical structure. After completing 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, and has good learnability and numerical convergence characteristics. Based on the above normalized feature vectors, a set of multi-parameter nonlinear regression equations is established to construct the inverse kinematic 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 set of equations is to output the corresponding control parameter values at a given three-dimensional coordinate input. Since the motion behavior of the probe station is inherently strongly nonlinear, state-coupled, and dynamically constrained, this nonlinear regression problem is embedded in a particle swarm-neural network hybrid solver for solution. This hybrid solver combines the global search ability of the particle swarm algorithm and the nonlinear fitting ability of the neural network, and adopts an alternating optimization strategy. In each iteration, the parameters of the equations are first preliminarily searched by the particle swarm algorithm, and then the neural network is used to perform gradient adjustment on the fitting residuals of each regression function, and the network weights and regression coefficients are dynamically updated, so as to ensure that the model can not only accurately approximate the sampling points but also has strong generalization ability. The constructed inverse kinematic model can automatically predict the corresponding probe control instructions after inputting three-dimensional coordinates, and has real-time response ability and high robustness. After the model is established, it is actually applied to the generation of control strategies. The three-dimensional position differences between each pair of adjacent test points are extracted from the globally optimal path. These differences reflect the physical movement behaviors required for the probe to complete the tests before and after.The displacement difference is input into the inverse kinematics model, and the model outputs the corresponding control instruction sequence. Each control quantity includes the change amplitude of the horizontal angle, the displacement step in the vertical direction, and the rotation step required for attitude adjustment. These control quantities together constitute the action plan that the probe needs to execute 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 the displacement control quantity sequence mainly includes two aspects: curve fitting and trajectory smoothing. Curve fitting uses interpolation methods such as cubic splines or Bezier curves to connect discrete control points into a continuous curve, ensuring that the probe movement path is continuously differentiable in geometric space, thereby avoiding speed mutations and excessive impact forces that may cause equipment accuracy loss or hardware wear. Trajectory smoothing constrains the acceleration change and angle change rate of the curve, making the trajectory meet the kinematic and dynamic constraints of the probe platform at the physical execution level, such as maximum speed, maximum acceleration, maximum rotation angle, etc. At the same time, intermediate transition points are inserted when necessary for dynamic attitude correction to improve trajectory stability and motion safety. The finally output probe trajectory control sequence has high spatial accuracy and takes into account the response characteristics of the mechanical structure and path continuity during the time execution process.

[0032] Referring to Figure 2 , this embodiment provides a wafer probe station detection path planning device, including: A boundary enhancement module 1 for performing boundary enhancement processing on the wafer surface image to obtain a boundary enhancement input tensor; A feature extraction module 2 for performing dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map containing the chip boundary and test point positions; An attention screening module 3 for performing channel attention and spatial attention screening on the first feature map to obtain a second feature map; A distribution feature analysis module 4 for performing chip distribution feature analysis on the second feature map to obtain a test point clustering result; A path planning module 5 for performing probe movement path planning based on the test point clustering result to obtain a globally optimal path; A calculation module 6 for establishing an inverse kinematics model of the probe station based on the globally optimal path and calculating a probe trajectory control sequence according to the inverse kinematics model of the probe station.

[0033] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to that described in the above method embodiment, and details are not described herein again.

[0034] Referring to Figure 3 , this embodiment of the present invention also provides a computer device, which may be a server, and its internal structure may be as Figure 3As shown in the figure. The computer device includes a processor, a memory, a display screen, an input device, a network interface, and a database connected through a system bus. Among them, the processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0035] Those skilled in the art can understand that Figure 3 the structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0036] An embodiment of the present invention further provides a probe station, which is used to implement the steps of any one of the above methods.

[0037] It should be noted that in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, device, article or method. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, device, article or method including the element.

[0038] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structural or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for detecting path planning of a wafer probe station, characterized in that, Including: Performing boundary enhancement processing on the wafer surface image to obtain a boundary-enhanced input tensor; Performing dual-path feature extraction based on the boundary-enhanced input tensor to obtain a first feature map including chip boundaries and test point positions; Performing channel attention and spatial attention screening on the first feature map to obtain a second feature map; Performing chip distribution feature analysis on the second feature map to obtain a test point clustering result; Performing probe movement path planning based on the test point clustering result to obtain a globally optimal path; Establishing an inverse kinematics model of the probe station based on the globally optimal path and calculating a probe trajectory control sequence according to the inverse kinematics model of the probe station.

2. The wafer probe station detection path planning method according to claim 1, wherein The performing boundary enhancement processing on the wafer surface image to obtain a boundary-enhanced input tensor includes: Inputting the wafer surface image into convolutional layers with three different-sized convolutional kernels respectively for feature extraction to obtain three groups of feature maps; Performing weighted fusion processing on the three groups of feature maps to obtain a fused feature map; Performing edge detection on the fused feature map using a boundary extraction operator, where the boundary extraction operator includes a horizontal Sobel operator, a vertical Sobel operator, and a Laplacian operator, to obtain a boundary feature map; Performing adaptive threshold segmentation on the boundary feature map to obtain a boundary-enhanced map, and performing channel splicing of the boundary-enhanced map and the wafer surface image to obtain a boundary-enhanced input tensor.

3. The wafer probe station detection path planning method according to claim 1, wherein The performing dual-path feature extraction based on the boundary-enhanced input tensor to obtain a first feature map including chip boundaries and test point positions includes: Inputting the boundary-enhanced input tensor into the encoder of the main path for multi-level feature extraction to obtain four encoded feature maps with different scales, where the encoder includes four downsampling modules; Performing decoding processing on the four encoded feature maps with different scales through the decoder of the main path to obtain a main path decoded feature map, where the decoder includes four upsampling modules; Selecting three feature maps at different levels from the main path encoded feature maps and inputting them into the supervision path to obtain three supervised branch feature maps; Performing channel dimensionality reduction and feature extraction on the three supervised branch feature maps respectively to obtain supervised feature maps, and fusing the supervised feature maps after size adjustment with the main path decoded feature map to obtain a fused feature representation; Performing semantic segmentation processing on the fused feature representation to obtain a first feature map including chip boundaries and test point positions.

4. The wafer probe station detection path planning method according to claim 1, wherein The performing channel attention and spatial attention screening on the first feature map to obtain a second feature map includes: Performing global average pooling and global max pooling operations on the first feature map to obtain channel dimension statistical features; Inputting the channel dimension statistical features into two fully connected layers respectively for dimensionality reduction and dimensionality increase processing to obtain a channel attention weight vector; Performing average pooling and max pooling on the first feature map along the channel dimension respectively to obtain two single-channel feature maps; Splicing the two single-channel feature maps and inputting them into a 7×7 convolutional layer to obtain a spatial attention weight map; Multiply the channel attention weight vector and the first feature map channel - by - channel, and then multiply the result element - by - element with the spatial attention weight map to obtain a preliminarily screened feature map; According to the position of the preliminarily screened feature map in the network, apply residual attention mechanisms with different coefficients to the shallow - layer feature map and the deep - layer feature map respectively to obtain a second feature map, where the residual attention mechanism is used to retain chip layout details and test - point distribution information.

5. The wafer probe station detection path planning method according to claim 1, wherein Performing chip distribution feature analysis on the second feature map to obtain a test - point clustering result, including: Performing semantic segmentation on the second feature map to obtain a binary test - point distribution map; Performing connected - component analysis on the binary test - point distribution map to obtain the basic features of the chip region and test points, where the basic features include the center coordinates, area, and the number of test points; Construct a chip - distribution covariance matrix based on the basic features, and apply a spectral clustering algorithm to the chip - distribution covariance matrix to obtain K chip clusters; Calculate and sort 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; Based on the test - priority region - division information, perform fine - grained clustering on the test points in the high - density region to obtain a test - point clustering result.

6. The wafer probe station detection path planning method according to claim 1, wherein Performing probe - movement path planning based on the test - point clustering result to obtain a globally optimal path, including: Extract the test - point density - distribution feature and spatial - distribution feature of each cluster based on the test - point clustering result to obtain a feature - vector matrix; Perform non - linear weighted calculation on the feature - vector matrix to obtain a path - planning objective function; Divide the test - point clustering result into multiple density - level regions, and assign test - priority weights to each density - level region to obtain a hierarchical detection sequence; Construct local sub - paths for the test points in each density - level region based on the hierarchical detection sequence to obtain a set of intra - region optimal paths; Perform multi - objective particle - swarm optimization on the set of intra - region optimal paths to obtain an inter - region path - connection scheme; Based on the inter - region path - connection scheme, perform adaptive annealing - tabu search hybrid optimization on the global path to obtain a globally optimal path.

7. The wafer probe station detection path planning method according to claim 1, characterized in that, Establishing an inverse kinematics model of the probe station based on the globally optimal path, and calculating a probe - trajectory control sequence according to the inverse kinematics model of the probe station, including: Extract a three - dimensional coordinate sequence of test points from the globally optimal path, and construct a probe - station motion mapping sample by combining the three - dimensional coordinate sequence with the corresponding three - axis control parameters of the probe station, where the three - axis control parameters of the probe station include the horizontal movement angle, vertical displacement, and rotation angle of the probe station; Perform feature standardization and spatial transformation on the probe - station motion mapping sample to obtain a normalized feature vector; A multi-parameter non-linear regression equation set is established based on the normalized eigenvector, and the multi-parameter non-linear regression equation set is embedded into a particle swarm-neural network hybrid solver to obtain a probe station inverse kinematics model. The multi-parameter non-linear regression equation set includes three groups of independent equations, and each group of independent equations corresponds to the control parameters of three degrees of freedom of the probe station. The particle swarm-neural network hybrid solver is used to alternately iterate and optimize the coefficients of the multi-parameter non-linear regression equation set and the neural network weights; Based on the difference in the positions of adjacent test points on the globally optimal path input by the probe station inverse kinematics model, a probe station displacement control quantity sequence is obtained. Each displacement control quantity in the probe station displacement control quantity sequence includes a single-step angle change amount, a displacement step length, and a rotation step length; Based on the probe station displacement control quantity sequence, curve fitting and trajectory smoothing processing are performed to obtain a probe trajectory control sequence.

8. A wafer probe station detection path planning device, characterized in that For implementing the steps of the wafer probe station detection path planning method according to any one of claims 1 to 7, the wafer probe station detection path planning device includes: A boundary enhancement module, configured to perform boundary enhancement processing on the wafer surface image to obtain a boundary enhancement input tensor; A feature extraction module, configured to perform dual-path feature extraction based on the boundary enhancement input tensor to obtain a first feature map including the chip boundary and the test point positions; An attention screening module, configured to perform channel attention and spatial attention screening on the first feature map to obtain a second feature map; A distribution feature analysis module, configured to perform chip distribution feature analysis on the second feature map to obtain a test point clustering result; A path planning module, configured to perform probe movement path planning based on the test point clustering result to obtain a globally optimal path; A calculation module, configured to establish a probe station inverse kinematics model based on the globally optimal path and calculate a probe trajectory control sequence according to the probe station inverse kinematics model.

9. A computer device, comprising a memory and a processor, wherein a computer program is stored in the memory, characterized in that, When the processor executes the computer program, the steps of the wafer probe station detection path planning method according to any one of claims 1 to 7 are implemented.

10. A probe station, characterized in that, The probe station is used to implement the steps of the wafer probe station detection path planning method according to any one of claims 1 to 7.

Images