A vision-based method and system for detecting and compensating wafer particle position offset

Abstract

This invention provides a vision-based method and system for detecting and compensating wafer particle position offset. It generates an initial imaging data set containing the surface microstructure by performing dual-view synchronous imaging on the wafer surface. Parallax matching and feature segmentation are then performed on the initial imaging data set to generate three-dimensional height information and an initial set of particle coordinates on the wafer surface. Based on the initial particle coordinate set and the charge distribution characteristics of the wafer surface, particle displacement is predicted and corrected, generating displacement compensation feature quantities for each particle. The fine-tuning optical system of the wafer inspection machine is invoked to perform optical correction on the displacement compensation feature quantities, generating corrected real-time imaging data. Coordinate registration is performed on the real-time imaging data to generate the final set of wafer particle position coordinates. This invention can improve the spatial dimensional integrity of wafer particle position detection, the comprehensiveness of multi-factor displacement prediction, and the actual physical accuracy of the detection results.

Description

Technical Field

[0001] This invention relates to the field of inspection, and more specifically, to a method and system for detecting and compensating for wafer particle position offset based on visual inspection. Background Technology

[0002] In the semiconductor manufacturing field, accurate detection of particle defects on the wafer surface is a crucial step in ensuring chip yield and reliability. Wafer particle position detection uses a single-view optical imaging system to acquire images of the wafer surface. After image segmentation and feature extraction, the planar coordinate information of the particles is obtained, and positional offsets during the detection process are compensated based on a robotic arm kinematic model. In the displacement compensation stage, traditional methods often only consider the physical offset caused by the robotic arm's movement, without fully considering the adsorption or repulsion effects of charge distribution on the wafer surface caused by processes such as friction and ion implantation on the particles. This leads to deviations in the prediction of the actual position offset of charged particles. At the same time, traditional detection methods rely heavily on post-processing data calibration to address the inherent distortion and focus drift of the optical system, lacking a real-time dynamic correction mechanism during the imaging process, which further affects the final accuracy of particle position detection. Summary of the Invention

[0003] This invention provides a method and system for detecting and compensating for wafer particle position offset based on visual inspection.

[0004] In a first aspect, embodiments of the present invention provide a method for detecting and compensating wafer particle position offset based on visual inspection. The method includes: performing dual-view synchronous imaging on the wafer surface to generate an initial imaging data set containing the surface microstructure, wherein the initial imaging data set consists of a first imaging sequence output by a first-view imaging unit and a second imaging sequence output by a second-view imaging unit, wherein the first imaging sequence and the second imaging sequence have the same timestamp and the imaging fields of view have a preset overlapping area; performing parallax matching and feature segmentation on the initial imaging data set to generate three-dimensional height information of the wafer surface particles and an initial particle coordinate set, wherein the initial particle coordinate set includes planar coordinate features of each particle in the imaging coordinate system and the corresponding height features; performing particle displacement prediction and correction based on the initial particle coordinate set and the charge distribution features of the wafer surface to generate displacement compensation features for each particle; calling the fine-tuning optical system of the wafer inspection machine to perform optical correction response on the displacement compensation features to generate corrected real-time imaging data; and performing coordinate registration on the real-time imaging data to generate a final set of wafer particle position coordinates.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory storing a computer program; and a processor for loading the computer program to implement the vision-based wafer particle position offset detection and compensation method as described above.

[0006] This invention provides a vision-based wafer particle position offset detection and compensation method. By constructing an initial imaging data set including timestamp synchronization and field-of-view overlap constraints through dual-view synchronous imaging, it overcomes the limitation of traditional single-view imaging, which can only acquire two-dimensional planar information. Through the collaborative processing of parallax matching and feature segmentation, the two-dimensional imaging information is transformed into a three-dimensional initial particle coordinate set containing planar coordinate features and height features. This solves the technical problem that traditional two-dimensional coordinates cannot reflect the spatial distribution characteristics of particles, extending particle position representation from a planar dimension to a three-dimensional spatial dimension. By introducing the fusion analysis of wafer surface charge distribution characteristics and the initial particle coordinate set, a particle position offset coupling mechanical motion and charge interaction is constructed. The particle displacement prediction model breaks through the limitation of traditional displacement compensation relying solely on mechanical kinematic models. It achieves comprehensive prediction of multi-physics field factors for particle position displacement. By fine-tuning the optical system to perform real-time optical correction response to displacement compensation feature quantities, it upgrades traditional post-processing data correction to hardware-level dynamic compensation during the imaging process. This eliminates the influence of inherent optical system distortion and focus drift on imaging quality. Finally, it generates the final set of wafer particle position coordinates through coordinate registration, significantly improving the spatial dimensional integrity of wafer particle position detection, the comprehensiveness of multi-factor displacement prediction, and the actual physical accuracy of the detection results. This provides more accurate position reference data for particle defect control in the wafer manufacturing process. Attached Figure Description

[0007] Figure 1 This is a flowchart of a wafer particle position offset detection and compensation method based on visual inspection provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0008] Please see Figure 1 This is a flowchart of a wafer particle position offset detection and compensation method based on visual inspection provided in an embodiment of the present invention. The method can be executed by a computer system and may include the following steps: Step S100: Perform dual-view synchronous imaging on the wafer surface to generate an initial imaging data set containing the surface microstructure. The initial imaging data set consists of a first imaging sequence output by a first-view imaging unit and a second imaging sequence output by a second-view imaging unit. The first and second imaging sequences have the same timestamp and their imaging fields of view have a preset overlapping area. Dual-view synchronous imaging refers to using two imaging units with different perspectives to simultaneously image the wafer surface, thus acquiring information about the wafer surface from different angles. The first and second perspective imaging units are devices used to acquire imaging data, such as cameras with set optical parameters and imaging performance. The first imaging sequence is a series of image data obtained by the first-view imaging unit continuously imaging the wafer surface. Similarly, the second imaging sequence is the imaging data sequence obtained by the second-view imaging unit. The timestamp is used to mark the shooting time of each imaging frame, ensuring that the first and second imaging sequences are synchronized in time, facilitating the analysis and processing of images from different perspectives at the same time. The preset overlapping area refers to the pre-defined overlapping portion in the imaging fields of view of the first and second perspective imaging units. In practice, for example, the first-view imaging unit and the second-view imaging unit can be arranged above the wafer at certain angles and positions, so that their imaging fields of view have a preset overlapping area. Then, through a synchronous triggering mechanism, the two imaging units simultaneously start to continuously image the wafer surface, generating a first imaging sequence and a second imaging sequence, and adding the same timestamp to each imaging frame, finally obtaining an initial imaging data set containing the surface microstructure; Step S200: Perform disparity matching and feature segmentation on the initial imaging data set to generate three-dimensional height information of particles on the wafer surface and an initial particle coordinate set. The initial particle coordinate set includes the planar coordinate features of each particle in the imaging coordinate system and the corresponding height features.

[0009] Parallax matching refers to the process of finding corresponding feature points in the first and second imaging sequences. Due to the different viewpoints, the position of the same object will differ in the two images. By calculating this difference (i.e., parallax), a basis can be provided for calculating the object's three-dimensional information. Feature segmentation is the operation of separating different features (such as particles on a wafer surface) from the background in an image to more accurately analyze the characteristics of the particles. Three-dimensional height information represents the height of the particles on the wafer surface relative to a reference plane, which helps to understand the spatial distribution of the particles. The initial particle coordinate set is a set containing the planar coordinate features of each particle in the imaging coordinate system (i.e., the position of the particle on the image plane) and the corresponding height features.

[0010] In performing disparity matching and feature segmentation, the first and second imaging sequences are processed to identify corresponding feature points and calculate disparity. Then, based on the disparity information and imaging system parameters, the three-dimensional height information of the particles is calculated using principles such as triangulation. Simultaneously, feature segmentation is performed on the image to determine the planar coordinates of each particle in the imaging coordinate system. For example, a matching algorithm based on feature point descriptors can be used for disparity matching, while feature segmentation can be performed using methods such as thresholding and region growing.

[0011] As one implementation, step S200 may include the following steps S210~S250: Step S210: Perform imaging consistency calibration processing on the first imaging sequence and the second imaging sequence in the initial imaging data set, adjust the imaging distortion parameters of the two imaging sequences based on the feature point correspondence of the preset overlapping area, and generate the first calibrated imaging sequence and the second calibrated imaging sequence after distortion calibration.

[0012] Imaging consistency calibration is performed to eliminate image distortion caused by factors such as the optical characteristics of the imaging unit, making the first and second imaging sequences more geometrically consistent and facilitating operations such as disparity matching. The feature point correspondence within the preset overlap region refers to the positional relationship between feature points of the same object in the first and second imaging sequences within the preset overlap region. Imaging distortion parameters describe the distortions caused by the imaging unit, such as radial distortion and tangential distortion.

[0013] In practice, feature points in the first and second imaging sequences are first identified within a pre-defined overlapping area, and their correspondences are determined using a feature point matching algorithm. Then, based on these correspondences, a calibration algorithm is used to adjust the imaging distortion parameters. For example, the Zhang Zhengyou calibration method can be used. By acquiring multiple images with known feature point locations, the distortion parameters of the imaging unit are calculated, and distortion correction is applied to the first and second imaging sequences to generate a first and second calibrated imaging sequence after distortion calibration.

[0014] Step S220: Calculate the gray-level co-occurrence matrix for the first and second calibration imaging sequences, and extract the texture feature distribution map of the two calibration imaging sequences in the overlapping area. The texture feature distribution map contains gray-level change frequency parameters of different directional channels.

[0015] The gray-level co-occurrence matrix (GLCM) is a matrix used to describe the texture features of an image, reflecting the spatial correlation between different gray levels. By calculating the GLCM, texture feature information of an image can be extracted. A texture feature distribution map is a visual representation of the texture features of two calibrated imaging sequences in the overlapping region. The gray-level change frequency parameters of different directional channels describe the frequency of gray-level value changes in different directions, and these parameters can reflect features such as texture roughness and directionality of the image.

[0016] When calculating the gray-level co-occurrence matrix (GLCM), the calculation parameters, such as distance and angle, are first determined. Then, based on the GLCM, the gray-level change frequency parameters for different directional channels, such as contrast, correlation, energy, and entropy, are calculated. Finally, these parameters are combined to form a texture feature distribution map. Specifically, the calculation can be performed by traversing each pixel within the overlapping region, statistically analyzing its gray-level combinations with neighboring pixels to construct the GLCM, and then calculating the corresponding texture feature parameters based on the matrix elements.

[0017] Step S230: Perform cross-view feature point matching based on texture feature distribution map, and use epipolar constraint algorithm to select matching feature point pairs that meet the disparity geometry conditions to generate an initial disparity candidate point set.

[0018] Cross-view feature point matching is the process of finding corresponding feature points between the first and second calibration imaging sequences. Using texture feature distribution maps can improve the accuracy of feature point matching. The epipolar constraint algorithm, based on binocular vision principles, utilizes the geometric relationship between two imaging units to limit the search range for feature point matching, improving matching efficiency and accuracy. Disparity geometry conditions refer to the geometric relationship that the disparity between corresponding feature points must satisfy in binocular vision. By filtering matching feature point pairs that satisfy the disparity geometry conditions, some incorrect matches can be eliminated. The initial disparity candidate point set is the set of possible matching feature point pairs obtained after filtering. Each element contains the coordinate information of the matching point pair and the corresponding disparity estimate.

[0019] In practice, feature points are first extracted from the first and second calibration imaging sequences based on the texture feature distribution map, and their feature descriptors are calculated. Then, a feature descriptor matching algorithm is used to find possible matching feature point pairs. Next, using the epipolar constraint algorithm, the epipolar trajectory corresponding to each feature point is calculated based on the intrinsic and extrinsic parameter matrices of the two imaging units. Matching feature point pairs located within a preset geometric range of the epipolar trajectory are selected. These point pairs are the matching feature point pairs that satisfy the disparity geometry conditions, and finally, an initial disparity candidate point set is generated.

[0020] As one implementation, step S230 may include the following steps S231 to S235: Step S231: Perform multi-scale feature extraction processing on the texture feature distribution map, construct a feature pyramid structure containing multiple scale levels, extract feature points with scale invariance at different scale levels, and generate a first feature point set and a second feature point set containing scale parameters and orientation parameters. The first feature point set corresponds to the first calibration imaging sequence, and the second feature point set corresponds to the second calibration imaging sequence.

[0021] Multi-scale feature extraction aims to analyze texture feature distribution maps at different scales to extract scale-invariant feature points. Scale invariance refers to the property that the feature descriptors of feature points remain unchanged when the image undergoes scale transformations (such as scaling up or down), thus improving the stability of feature point matching. A feature pyramid structure is a hierarchical structure that represents an image at different scales, with each scale level corresponding to a different image resolution. By constructing a feature pyramid, feature points can be extracted at different scales. The scale parameter describes the scale level of the feature point, and the orientation parameter represents the principal direction of the feature point.

[0022] When constructing the feature pyramid, the texture feature distribution map is first Gaussian smoothed to obtain smoothed images at different scales. These smoothed images are then arranged in descending or ascending order of scale to form the feature pyramid structure. At each scale level, a feature point detection algorithm, such as SIFT (Scale Invariant Feature Transform), is used to extract scale-invariant feature points. For each feature point, its scale and orientation parameters are calculated. For example, the SIFT algorithm identifies feature points by finding local extrema in Gaussian difference images at different scales, and then calculates its orientation parameter based on the gradient information of the feature point's neighborhood. Finally, a first feature point set corresponding to the first calibration imaging sequence and a second feature point set corresponding to the second calibration imaging sequence are obtained.

[0023] Step S232: Calculate the feature descriptors of the first feature point set and the second feature point set. Use a dynamic weighted similarity fusion algorithm to measure the similarity between the feature descriptors of the first feature point set and the feature descriptors of the second feature point set, and generate an initial feature matching matrix. The row dimension of the initial feature matching matrix corresponds to the number of feature points in the first feature point set, and the column dimension corresponds to the number of feature points in the second feature point set.

[0024] Feature descriptors are vectors used to describe the features of feature points, containing local image information such as gradient direction and grayscale value. By calculating feature descriptors, the features of feature points can be quantified, facilitating matching operations. The dynamic weighted fusion similarity algorithm is used to calculate the similarity between two feature descriptors. By dynamically adjusting the weights of different feature dimensions, it more accurately measures the similarity between feature descriptors. The initial feature matching matrix is ​​a two-dimensional matrix used to record the matching results between the first and second feature point sets. Each element in the matrix represents the similarity between the corresponding feature points.

[0025] When calculating feature descriptors, the same algorithm used for feature point extraction, such as the SIFT algorithm, can be used to generate corresponding feature descriptors for each feature point in the first and second feature point sets. Then, a dynamic weighted similarity fusion algorithm is used to calculate the similarity between each feature descriptor in the first and second feature point sets. For example, feature descriptors can be viewed as vectors in a high-dimensional space, and similarity can be measured by calculating the distance between vectors (such as Euclidean distance, cosine distance, etc.). During dynamic weighted fusion, the weight of each dimension can be dynamically adjusted based on its importance to the matching. Finally, the calculated similarity values ​​are filled into the initial feature matching matrix, where the row dimension corresponds to the number of feature points in the first feature point set, and the column dimension corresponds to the number of feature points in the second feature point set.

[0026] Step S233: Construct an epipolar geometric model based on the intrinsic and extrinsic parameter matrices of the dual-view imaging system, calculate the epipolar trajectory corresponding to each first feature point according to the epipolar geometric model, and screen candidate matching point pairs located within the preset geometric range of the epipolar trajectory in the initial feature matching matrix.

[0027] The intrinsic parameter matrix of a dual-view imaging system describes the internal optical characteristics of the imaging units, such as focal length and principal point position, while the extrinsic parameter matrix describes the relative position and attitude relationship between the two imaging units. The epipolar geometry model is a model built based on the geometric relationships of the dual-view imaging system, describing the epipolar line corresponding to a point on one imaging plane on another. The epipolar trajectory refers to the position of the epipolar line corresponding to a certain feature point in the first calibration imaging sequence within the second calibration imaging sequence. The preset geometric range is a pre-defined area used to filter candidate matching point pairs located near the epipolar line. When constructing the epipolar geometry model, the intrinsic and extrinsic parameter matrices of the dual-view imaging system are first obtained, which can be obtained through calibration and other methods. Then, based on the principles of epipolar geometry, the epipolar equation corresponding to each first feature point in the second calibration imaging sequence is calculated using the intrinsic and extrinsic parameter matrices. Next, the initial feature matching matrix is ​​traversed. For each feature point in the first feature point set, its matching points in the second feature point set are selected, and the positions of these matching points must be within the preset geometric range of the epipolar trajectory corresponding to that feature point. These matching point pairs are the candidate matching point pairs.

[0028] Step S234: Perform bidirectional consistency verification on the candidate matching point pairs, and retain the mutually inverse matching point pairs that simultaneously satisfy the condition that both the path from the first feature point to the second feature point and the path from the second feature point to the first feature point are optimal matches.

[0029] Bidirectional consistency verification is performed to ensure the accuracy of matched point pairs and avoid one-way matching errors. A mutually inverse matching point pair is a point pair that is the optimal match both from the first feature point to the second feature point and from the second feature point to the first feature point.

[0030] During bidirectional consistency verification, for each candidate matching point pair, firstly, from the perspective of the first feature point set, it is checked whether the matching point pair is the optimal match of the first feature point in the second feature point set. Then, from the perspective of the second feature point set, it is checked whether the matching point pair is the optimal match of the second feature point in the first feature point set. Only matching point pairs that satisfy both conditions are retained as inverse matching point pairs. For example, the optimal matching point can be found by comparing the similarity value of each feature point with other matching points. If a matching point pair is the optimal match in both directions, it is considered to have passed the bidirectional consistency verification.

[0031] Step S235: Fuse the scale and orientation parameters of the inverse matching point pairs to generate an initial disparity candidate point set with geometric constraints and scale consistency. Each element in the initial disparity candidate point set contains the coordinate information of the matching point pair and the corresponding disparity estimate.

[0032] The fusion of scale and orientation parameters of inverse matching point pairs aims to integrate the scale and orientation information of the matching point pairs, giving the elements in the initial disparity candidate point set more comprehensive feature information. Geometric constraints refer to using the geometric relationships of the dual-view imaging system to constrain the matching point pairs, ensuring the accuracy of disparity estimation. Scale consistency means that the scale parameters of the matching point pairs should have a certain degree of consistency under different viewpoints, which can improve the reliability of disparity estimation. The disparity estimate refers to the positional difference of the matching point pair in two imaging sequences; calculating the disparity estimate provides a basis for calculating 3D height information.

[0033] When fusing scale and orientation parameters, the scale and orientation parameters of inverse matching point pairs can be weighted and averaged or fused using other methods to obtain the fused scale and orientation parameters. Then, based on the coordinate information of the matching point pairs, the disparity estimate between them is calculated. The fused scale parameters, orientation parameters, coordinate information of the matching point pairs, and disparity estimate are combined into a single element to finally generate an initial set of disparity candidate points with geometric constraints and scale consistency.

[0034] Step S240: Outlier removal is performed on the initial disparity candidate point set. Inner point feature point pairs are iteratively selected using the random sampling consensus algorithm. The disparity values ​​of each inner point feature point pair are calculated and a dense disparity map is generated.

[0035] Outlier removal is performed to eliminate potentially mismatched point pairs from the initial disparity candidate point set. These mismatched point pairs may be caused by factors such as noise or changes in illumination, affecting the accuracy of disparity calculation. Random Sample Consensus (RANSAC) is an iterative algorithm used to estimate mathematical model parameters. It selects a subset of point pairs from the initial disparity candidate point set through random sampling to fit the model, and then evaluates other point pairs based on the model, filtering out inlier feature point pairs that conform to the model. Inlier feature point pairs are those that conform to the fitted model and are relatively reliable matching pairs. A dense disparity map is an image representing disparity information, where the gray value of each pixel represents the disparity value at that location.

[0036] When using the random sample consensus algorithm, a subset of initial disparity candidate point pairs is first randomly selected, and a disparity model (such as a planar model) is fitted based on these point pairs. Then, the distances from other point pairs to this model are calculated, and point pairs with distances less than a preset threshold are marked as inlier feature point pairs. This process is repeated multiple times, each time randomly selecting different point pairs for fitting, and finally, the model with the most inlier feature point pairs is selected as the optimal model. Finally, the disparity values ​​of all inlier feature point pairs are calculated, and these disparity values ​​are filled into the corresponding pixel positions to generate a dense disparity map.

[0037] Step S250: Combining the dense parallax map and the preset baseline parameters of the dual-view imaging system, the three-dimensional height information of the particles on the wafer surface is generated by converting the data through the principle of triangulation. The first calibration imaging sequence is then processed by region growth segmentation to extract the boundary contour features of each particle region and generate an initial set of particle coordinates containing planar coordinate features and height features.

[0038] Baseline parameters in a dual-view imaging system refer to geometric parameters such as the distance between two imaging units. Triangulation is based on the principle of similar triangles, using disparity information and baseline parameters to calculate the three-dimensional position of an object. Region growing segmentation is an image segmentation method that starts from one or more seed points and merges adjacent pixels with similar features into a single region, thereby segmenting the image into different regions. Boundary contour features refer to the boundary shape information of particle regions; extracting boundary contour features can more accurately determine the position and shape of particles. Planar coordinate features represent the planar position of particles in the imaging coordinate system, while height features represent the three-dimensional height information of particles.

[0039] When calculating 3D height information, based on the principle of triangulation, combined with the disparity values ​​in the dense disparity map and the preset baseline parameters of the dual-view imaging system, the 3D height value corresponding to each pixel is calculated, thus obtaining the 3D height information of the particles on the wafer surface. During region growth and segmentation, a suitable seed point is first selected, which can be chosen based on features such as image grayscale values ​​and texture. Then, starting from the seed point, adjacent pixels with similar grayscale values ​​or other features are merged into the same region until certain stopping conditions are met, such as region size and grayscale difference. Next, edge detection and other algorithms are used to extract the boundary contour features of each particle region. Finally, the planar coordinate features of the particle region (such as centroid coordinates) are associated with the corresponding height features to generate an initial set of particle coordinates containing both planar coordinate features and height features.

[0040] As one implementation method, in step S250, the first calibration imaging sequence is subjected to region growing segmentation processing to extract the boundary contour features of each particle region and generate an initial particle coordinate set containing planar coordinate features and height features. Specifically, it may include the following steps S251~S255: Step S251: Adaptive threshold segmentation is performed on the first calibration imaging sequence. The dynamic segmentation threshold is calculated based on the local gray mean and standard deviation to generate a binary imaging mask. In the binary imaging mask, the particle region and the background region have different gray value distributions.

[0041] Adaptive thresholding is a segmentation method that automatically adjusts the threshold based on local image features, better adapting to grayscale variations in different regions of an image. The local grayscale mean refers to the average grayscale value of pixels within a specific local region of an image, while the standard deviation reflects the dispersion of grayscale values ​​within that region. Dynamic segmentation thresholding is calculated based on the local grayscale mean and standard deviation, and it varies depending on the local region of the image. Binarization imaging masks are images obtained by binarizing the image, where granular regions and background regions are represented by different grayscale values ​​(e.g., 0 and 255), facilitating processing and analysis.

[0042] In adaptive thresholding, the first calibration imaging sequence is first divided into multiple local regions, which can be done using fixed-size windows. For each local region, its gray-level mean and standard deviation are calculated. Then, according to a preset formula, the dynamic segmentation threshold for that region is calculated using the local gray-level mean and standard deviation. Next, the pixel gray-level values ​​within that region are compared with the dynamic segmentation threshold. Pixels with values ​​greater than the threshold are assigned one gray-level value (e.g., 255), and pixels with values ​​less than the threshold are assigned another gray-level value (e.g., 0), ultimately generating a binary imaging mask. For example, the dynamic segmentation threshold = local gray-level mean + k * standard deviation (where k is a preset coefficient).

[0043] Step S252: In the binary imaging mask, connected regions are marked using a multi-neighborhood region growing algorithm. Using gray value as the seed point selection condition, adjacent pixels that meet the gray-level similarity threshold are merged into the same connected region to generate a preliminary particle region marking map.

[0044] The multi-neighborhood region growing algorithm is an image segmentation algorithm based on region growing. It searches for adjacent pixels within multiple neighborhoods and merges pixels with similar features into the same region. The seed point is the starting point for region growing; in this step, grayscale value is used as the selection criterion, choosing pixels with a set grayscale value as seed points. The grayscale similarity threshold is used to determine whether adjacent pixels belong to the same connected region. If the difference between the grayscale value of an adjacent pixel and the grayscale value of the current region is less than the threshold, it is merged into that region. A connected region is a region composed of adjacent pixels with similar features. The preliminary particle region labeling map is the image obtained after labeling the connected regions in the binarized imaging mask, with each connected region represented by a different label value.

[0045] When labeling connected regions, the binary imaging mask is first traversed, and seed points are selected based on grayscale values. Then, starting from a seed point, neighboring pixels are searched within its multi-neighborhood, and the difference between the grayscale value of the neighboring pixel and the current region's grayscale value is calculated. If the difference is less than a grayscale similarity threshold, the neighboring pixel is merged into the current region, and the search continues within its neighborhood. This process is repeated until no neighboring pixels meet the criteria, forming a connected region. Next, the next seed point is selected, and the above process is repeated until the entire binary imaging mask has been traversed, ultimately generating a preliminary particle region labeling map.

[0046] Step S253: Combine the height feature value in the dense parallax map to perform region screening on the preliminary particle region labeling map, remove noise regions with height feature values ​​lower than the preset height threshold and tiny regions with areas smaller than the preset area threshold, and retain effective particle regions.

[0047] Height features refer to the 3D height information of each pixel in a dense parallax map. A preset height threshold is a pre-defined height value used to filter out granular regions with a certain height. Noise regions refer to false regions in the image caused by noise or other factors; their height features are usually low. A preset area threshold is a pre-defined area value used to filter out regions of a certain size. Tiny regions with an area smaller than this threshold may be caused by noise or other interference factors and need to be removed. Effective granular regions are those granular regions with a certain height and area that are retained after filtering.

[0048] During region filtering, each connected region in the initial particle region labeling map is first traversed to obtain the height feature of the pixels within that region in the dense disparity map. The average height feature of the region is calculated; if the average height feature is lower than a preset height threshold, the region is marked as a noise region and removed. Then, the area of ​​each connected region is calculated; if the area is less than a preset area threshold, the region is removed. Finally, the remaining regions are the effective particle regions.

[0049] Step S254: Extract the boundary contour of the effective particle region, obtain the closed contour curve of each particle region through the chain code tracking algorithm, and calculate the minimum bounding rectangle parameter and centroid coordinates of the contour curve as planar coordinate feature quantities.

[0050] Boundary contour extraction is used to determine the boundary shape of the effective particle region, facilitating the analysis of particle position and shape. Chain code tracking is an algorithm used to extract the boundary contours of objects in an image, describing the boundary shape by recording the connection directions of boundary pixels. A closed contour curve refers to the closed curve formed by the boundary of the particle region. The minimum bounding rectangle parameters refer to the parameters of the smallest rectangle that can completely contain the boundary contour curve of the particle region, such as the rectangle's position, size, and angle. Centroid coordinates refer to the coordinates of the geometric center of the particle region, which can be used as a representation of the particle's planar position in the imaging coordinate system.

[0051] During boundary contour extraction, for each effective particle region, a chain code tracing algorithm is used to trace the connection direction of boundary pixels starting from a boundary pixel in that region, following certain rules (such as clockwise or counterclockwise). The chain code value of each pixel is recorded, ultimately forming a closed contour curve. Then, based on the coordinate information of the closed contour curve, its minimum bounding rectangle parameter is calculated, which can be done using methods such as rectangle rotation. Simultaneously, the centroid coordinates of the region enclosed by the closed contour curve are calculated, and the minimum bounding rectangle parameter and centroid coordinates are used as planar coordinate feature quantities.

[0052] Step S255: Associate the centroid coordinates of the effective particle region with the height feature of the corresponding position in the dense parallax map to generate an initial particle coordinate set containing planar coordinate features and height features. The planar coordinate features and height features of each particle in the initial particle coordinate set are associated through the same particle identifier.

[0053] Associating centroid coordinates and height features is to map the planar position and three-dimensional height information of a particle in the imaging coordinate system, forming a complete particle coordinate information. A particle identifier is a unique number or label used to identify each particle, through which planar coordinate features and height features can be associated, facilitating management and querying.

[0054] During the association process, for each valid particle region, its centroid coordinates are obtained. Then, based on the centroid coordinates, the corresponding height feature is found in the dense disparity map. A unique particle identifier is assigned to this particle, and the centroid coordinates (as a planar coordinate feature) and the height feature are associated through this particle identifier. Finally, the planar coordinate features and height features of all particles are organized according to the particle identifiers to generate an initial set of particle coordinates containing both planar coordinate features and height features.

[0055] Step S300: Based on the initial set of particle coordinates and the characteristics of charge distribution on the wafer surface, predict and correct the particle displacement, and generate the displacement compensation feature quantity of each particle. The displacement compensation feature quantity is used to characterize the positional offset characteristics of the particle caused by the movement of the robotic arm.

[0056] Particle displacement prediction and correction are designed to account for the impact of factors such as robotic arm movement on the position of particles on the wafer surface. By predicting and correcting particle displacement, the accuracy of particle position detection is improved. Wafer surface charge distribution characteristics refer to information such as charge density and polarity in different regions of the wafer surface; these characteristics may affect particle movement and position. Displacement compensation features are used to compensate for positional shifts in particles caused by factors such as robotic arm movement, and include displacement components in the X and Y axes.

[0057] When predicting and correcting particle displacement, the relationship between the initial particle coordinate set and the wafer surface charge distribution characteristics is first analyzed. Based on this relationship, a suitable model or algorithm is used to predict the particle displacement during the robotic arm's movement. Then, displacement compensation features are generated based on the prediction results to correct the particle position. For example, a machine learning-based prediction model can be established, taking the initial particle coordinate set and wafer surface charge distribution characteristics as input and the particle displacement as output, and the model can be trained to predict particle displacement.

[0058] As one implementation, step S300 may include the following steps S310~S350: Step S310: Call the wafer charge detection unit to collect the charge distribution characteristics of the wafer surface. The charge distribution characteristics include the charge density distribution spectrum and charge polarity distribution matrix of each region on the wafer surface. The spatial resolution of the charge density distribution spectrum is consistent with the resolution of the planar coordinate feature quantity of the initial particle coordinate set.

[0059] A wafer charge detection unit is a device used to detect the charge distribution on the wafer surface. It acquires charge information from the wafer surface through monitoring principles such as electrostatic induction, using devices like Keithley electrometers, Trek electrostatic voltmeters, XPS charge analysis systems, and scanning capacitance microscopes (SCM). A charge density distribution map is an image representing the charge density of different regions on the wafer surface, where the grayscale value or color of each pixel represents the characteristic charge density value at that location. The charge polarity distribution matrix is ​​a matrix where each element represents the charge polarity at the corresponding location, such as positive, negative, or zero. Maintaining the spatial resolution of the charge density distribution map consistent with the resolution of the planar coordinate features of the initial particle coordinate set facilitates the correlation and analysis of charge distribution features with particle coordinate information.

[0060] When acquiring charge distribution characteristics, the wafer charge detection unit is controlled to scan and detect the wafer surface. For example, a line-by-line scan or a spiral scan can be used to detect the charge in each region of the wafer surface. The detected charge information is processed and transformed to generate a charge density distribution map and a charge polarity distribution matrix. When generating the charge density distribution map, the detected charge density data is mapped to the pixel values ​​of the image to ensure that the spatial resolution of the map is consistent with the resolution of the planar coordinate features of the initial particle coordinate set. For the charge polarity distribution matrix, the elements at corresponding positions are marked with the corresponding polarity identifiers based on the positive or negative value of the charge density.

[0061] As one implementation, step S310 may include the following steps S311 to S315: Step S311: Control the probe array of the wafer charge detection unit to perform gridded scanning on the wafer surface, the scanning path moves spirally along the radial direction of the wafer, and the scanning step size matches the resolution of the planar coordinate feature quantity of the initial particle coordinate set.

[0062] A probe array is a component in a wafer charge detection unit used to detect charge. It consists of multiple probes and can simultaneously detect charge information at multiple locations. Raster scanning divides the wafer surface into multiple small grid regions, which the probe array scans sequentially. Spiral movement refers to the scanning path moving in a spiral shape along the radial direction of the wafer, allowing for more comprehensive coverage of the wafer surface. The scan step size refers to the distance the probe array moves in each step during the scan. Matching this step size to the resolution of the planar coordinate features of the initial particle coordinate set ensures spatial consistency between the acquired charge information and the particle coordinate information.

[0063] During rasterized scanning, the scan step size is first determined based on the resolution of the planar coordinate features of the initial particle coordinate set. Then, the probe array is controlled to move along a spiral path across the wafer surface, starting from the center or edge, with each movement being the scan step size. At each scan position, the probe array acquires the charge-induced signal at that location.

[0064] Step S312: Collect charge induction signals at each scanning sampling point, amplify the weak induction signals through a charge amplifier, convert them into digital signals through an analog-to-digital converter, and generate an original charge induction data set. The original charge induction data set includes the charge value of each sampling point and the corresponding detection coordinate system coordinates.

[0065] The charge-induced signal is a weak charge signal detected by the probe array at each scanning sampling point. Because this signal is typically weak, it requires amplification. A charge amplifier is a device used to amplify the charge-induced signal, increasing its amplitude to a suitable level. An analog-to-digital converter (ADC) converts the analog signal into a digital signal, facilitating processing and analysis. The raw charge-induced data set is a dataset combining the charge value collected at each scanning sampling point with the corresponding coordinates in the detection coordinate system. The charge value represents the amount of charge detected at that sampling point, and the coordinates indicate the location of the sampling point within the detection coordinate system. During charge-induced signal acquisition, the probe array transmits the detected signal to the charge amplifier at each scanning sampling point. The charge amplifier amplifies the weak signal to a suitable amplitude. Then, the ADC converts the amplified analog signal into a digital signal, recording the charge value at that sampling point. Simultaneously, it records the coordinates of the sampling point in the detection coordinate system, generating the raw charge-induced data set.

[0066] Step S313: Noise suppression is performed on the original charge sensing data set. High-frequency noise is filtered out from the charge values ​​of each sampling point by using a wavelet threshold denoising algorithm, while retaining the low-frequency charge distribution characteristic components.

[0067] Noise suppression aims to remove potential noise interference from the original charge-induced data set, improving data quality and reliability. Wavelet thresholding denoising algorithms, based on wavelet transform, decompose the signal into components of different frequencies. By setting an appropriate threshold, high-frequency noise components are removed, while low-frequency useful signal components are retained. High-frequency noise refers to the higher-frequency noise components in the signal, typically caused by noise from the detection equipment, external interference, etc. Low-frequency charge distribution characteristic components refer to the low-frequency signal components that reflect the main characteristics of charge distribution on the wafer surface.

[0068] During noise suppression, the charge values ​​at each sampling point in the original charge-induced data set are first subjected to wavelet transform, decomposing them into wavelet coefficients of different frequencies. Then, a threshold is determined for the wavelet coefficients according to a preset threshold rule. Wavelet coefficients with absolute values ​​less than the threshold are set to zero, indicating that these coefficients correspond to high-frequency noise components; wavelet coefficients with absolute values ​​greater than the threshold are retained. Finally, an inverse wavelet transform is performed on the processed wavelet coefficients to obtain the denoised charge values, preserving the low-frequency charge distribution characteristic components.

[0069] Step S314: Based on the detection coordinates of the original charge sensing data set, spatial interpolation processing is performed on the denoised charge sensing data using a high-order interpolation algorithm to generate a charge density distribution map with continuous spatial distribution. The pixel values ​​of the charge density distribution map represent the charge density feature values ​​at the corresponding locations.

[0070] Higher-order interpolation algorithms are used to perform interpolation calculations between discrete data points, estimating data values ​​at unknown locations based on known discrete data points. Spatial interpolation is used to convert discrete, denoised charge-induced data into a charge density distribution map with a continuous spatial distribution. The pixel values ​​in the charge density distribution map represent the charge density feature values ​​at corresponding locations. Through spatial interpolation, the charge density distribution across the entire wafer surface can be obtained. During spatial interpolation, the positions of known data points are first determined based on the detection coordinates of the original charge-induced data set. Then, higher-order interpolation algorithms, such as cubic spline interpolation, are used to estimate the charge density feature values ​​at other locations based on the charge values ​​of these known data points. The estimated charge density feature values ​​are then filled into the corresponding pixel positions, generating a charge density distribution map with a continuous spatial distribution.

[0071] Step S315: Generate a charge polarity distribution matrix based on the charge density feature value sign of each pixel in the charge density distribution map. Pixels with positive charge density feature values ​​are marked as first polarity identifiers, pixels with negative charge density feature values ​​are marked as second polarity identifiers, and pixels with zero charge density feature values ​​are marked as third polarity identifiers, thus generating a charge polarity distribution matrix containing charge polarity information.

[0072] The charge polarity distribution matrix is ​​used to represent the charge polarity at various locations on the wafer surface. By judging the sign of the charge density characteristic value of each pixel in the charge density distribution map, its charge polarity can be determined. The first polarity identifier, the second polarity identifier, and the third polarity identifier are pre-defined identifiers used to represent different charge polarities. For example, 1 can be used to represent positive charge polarity (first polarity identifier), -1 can be used to represent negative charge polarity (second polarity identifier), and 0 can be used to represent zero charge polarity (third polarity identifier).

[0073] When generating the charge polarity distribution matrix, each pixel in the charge density distribution map is traversed to obtain its charge density feature value. Based on the sign of the charge density feature value, the pixel is marked with the corresponding polarity identifier. The polarity identifiers of all pixels are combined to generate a charge polarity distribution matrix containing charge polarity information.

[0074] Step S320: Perform coordinate transformation on the planar coordinate features of each particle in the initial particle coordinate set. Based on the calibration parameters of the dual-view imaging system, convert the planar coordinate features of the pixel unit in the imaging coordinate system into physical length units, and then map them to the charge detection coordinate system of the wafer charge detection unit to generate a particle mapping coordinate set in the charge coordinate system.

[0075] Coordinate transformation is performed to convert the planar coordinate features in the initial particle coordinate set from the imaging coordinate system to the charge detection coordinate system, so as to correlate the particle coordinate information with the charge distribution characteristics of the wafer surface. The calibration parameters of the dual-view imaging system describe the internal and external geometric parameters of the imaging system, enabling the conversion between the imaging coordinate system and the physical coordinate system. The pixel-unit planar coordinate feature represents the particle planar position in pixels within the imaging coordinate system, while the physical length unit is a more practically meaningful unit of length, such as millimeters. The charge detection coordinate system is the coordinate system used by the wafer charge detection unit, and the particle mapping coordinate set is the coordinate set obtained after transforming the particle planar coordinate features to the charge detection coordinate system.

[0076] During coordinate transformation, the pixel unit planar coordinate features in the imaging coordinate system are first converted into physical length units based on the calibration parameters of the dual-view imaging system. This can be calculated using scaling factors, translation parameters, etc., in the calibration parameters. Then, based on the relationship between the charge detection coordinate system and the physical coordinate system, the transformed physical coordinates are mapped to the charge detection coordinate system, ultimately generating a set of particle-mapped coordinates in the charge coordinate system.

[0077] Step S330: Based on the particle mapping coordinate set, query the charge density distribution map and charge polarity distribution matrix, extract the charge density feature value and charge polarity feature value of the region where each particle is located, and generate a particle charge correlation feature set. Each element in the particle charge correlation feature set corresponds one-to-one with the particles in the initial particle coordinate set.

[0078] Querying the charge density distribution map and charge polarity distribution matrix is ​​to obtain charge-related information about the region where each particle is located. The particle charge association feature set is a dataset that associates each particle with its corresponding charge density feature value and charge polarity feature value. Each element corresponds to a particle in the initial particle coordinate set, containing the charge density feature value and charge polarity feature value of the region where that particle is located.

[0079] During the query, each particle coordinate in the particle mapping coordinate set is traversed, and the corresponding pixel is found in the charge density distribution map and charge polarity distribution matrix based on that coordinate. The charge density feature value and charge polarity feature value of that pixel are obtained and associated with the particle. The charge-related information of all particles is combined to generate a particle charge association feature set.

[0080] Step S340: Call the pre-trained particle displacement prediction model to perform joint prediction processing on the height feature quantity and particle charge correlation feature set in the initial particle coordinate set, and generate the preliminary displacement prediction vector of each particle. The preliminary displacement prediction vector includes the displacement component in the X-axis direction and the displacement component in the Y-axis direction.

[0081] The pre-trained particle displacement prediction model is a model pre-trained to predict particle displacement. It learns the relationship between height features and particle charge correlation features from the initial particle coordinate set and particle displacement. The joint prediction process takes the height features and particle charge correlation features as input, considering their influence on particle displacement, and then uses the model to make predictions. The initial displacement prediction vector is the model's output prediction result, containing the particle displacement components in the X and Y axis directions.

[0082] In the joint prediction process, the height features and particle charge correlation features from the initial particle coordinate set are used as input data and fed into a pre-trained particle displacement prediction model. The model processes and analyzes the input data based on its internal structure and learned parameters, ultimately generating a preliminary displacement prediction vector for each particle. For example, the particle displacement prediction model can employ a neural network model, such as a multilayer perceptron (MLP), which includes an input layer, multiple hidden layers, and an output layer. The input layer receives the height features and particle charge correlation features; after nonlinear transformation by the hidden layers, the output layer outputs a preliminary displacement prediction vector, where each neuron corresponds to a displacement component along the X-axis or Y-axis.

[0083] As one implementation, step S340 may include the following steps S341 to S345: Step S341: Normalize the height feature quantity in the initial particle coordinate set, map the height feature quantity to a preset numerical range, and generate normalized height feature value.

[0084] Normalization is performed to unify the numerical range of height features within a preset interval, preventing the negative impact of numerical differences between different features on model training and prediction. The preset numerical interval is a pre-defined range, such as [0,1]. Normalized height feature values ​​are the height feature values ​​obtained after normalization. A min-max normalization method can be used during normalization. First, the minimum and maximum values ​​of the height features in the initial particle coordinate set are found. Then, according to the formula: Normalized height feature value = (Height feature value - Minimum value) / (Maximum value - Minimum value), each height feature is mapped to the preset numerical interval to generate a normalized height feature value.

[0085] Step S342: Standardize the charge density feature values ​​in the particle charge correlation feature set to eliminate the dimensional differences between different charge density feature values ​​and generate standardized charge density feature values.

[0086] Standardization is used to eliminate dimensional differences between different charge density eigenvalues, giving them the same scale and distribution, making it easier for models to process and analyze these features. Standardized charge density eigenvalues ​​are charge density eigenvalues ​​obtained after standardization.

[0087] When performing standardization, the Z-score standardization method can be used. Calculate the mean and standard deviation of the charge density eigenvalues ​​in the particle charge correlation feature set. Then, according to the formula: Standardized charge density eigenvalue = (Charge density eigenvalue - Mean) / Standard deviation, standardize each charge density eigenvalue to generate a standardized charge density eigenvalue.

[0088] Step S343: Perform feature concatenation on the normalized height feature value, standardized charge density feature value, and charge polarity feature value to generate a joint feature vector of particles containing multi-dimensional feature components. The dimension of the joint feature vector of particles matches the dimension of the input layer of the particle displacement prediction model.

[0089] Feature concatenation combines different types of feature values ​​to form a vector containing multi-dimensional feature components. The joint feature vector of particles contains feature components of multiple dimensions, such as normalized height feature values, normalized charge density feature values, and charge polarity feature values. Its dimensions need to match the input layer dimensions of the particle displacement prediction model so that it can be correctly input into the model for prediction.

[0090] During feature concatenation, the normalized height eigenvalues, normalized charge density eigenvalues, and charge polarity eigenvalues ​​are arranged in a specific order and combined into a vector, which is the particle joint feature vector. For example, if there are n normalized height eigenvalues, m normalized charge density eigenvalues, and k charge polarity eigenvalues, then the dimension of the particle joint feature vector is n+m+k.

[0091] Step S344: Input the joint feature vector of particles into the input layer of the particle displacement prediction model, and perform feature transformation processing through multiple fully connected hidden layers in sequence. Each fully connected hidden layer performs nonlinear mapping through a linear rectified activation function to generate a high-level abstract feature vector.

[0092] Fully connected hidden layers are intermediate layers in neural network models, where each neuron is connected to all neurons in the previous layer. Feature transformation processing refers to transforming and processing the input granular joint feature vector through fully connected hidden layers to extract higher-level feature information. The Rectified Linear Activation Function (ReLU) is a feasible activation function that can introduce non-linearity to enhance the model's expressive power. High-level abstract feature vectors are feature vectors with a higher level of abstraction obtained after processing through multiple fully connected hidden layers.

[0093] During feature transformation, the joint feature vector of particles is input into the input layer of the particle displacement prediction model. The input layer passes the data to the first fully connected hidden layer, where each neuron performs a weighted sum of the input data and then performs a nonlinear mapping through a linear rectified activation function to obtain the output of the hidden layer. This output is then used as the input to the next fully connected hidden layer, and so on, until a high-level abstract feature vector is finally generated.

[0094] Step S345: Input the high-level abstract feature vector into the output layer of the particle displacement prediction model, and output a preliminary displacement prediction vector containing the displacement components in the X-axis direction and the Y-axis direction through a linear activation function. Each component of the preliminary displacement prediction vector represents the predicted offset of the particle in the corresponding direction.

[0095] Linear activation functions directly use the input value as the output value without introducing nonlinear transformations. The output layer is the last layer of the particle displacement prediction model, outputting a preliminary displacement prediction vector based on the high-level abstract feature vector. Each component of the preliminary displacement prediction vector represents the predicted offset of the particle in the X and Y axis directions.

[0096] During output, the high-level abstract feature vector is input into the output layer of the particle displacement prediction model. The neurons in the output layer perform a weighted summation of the high-level abstract feature vector and output the result through a linear activation function, forming a preliminary displacement prediction vector containing displacement components in the X-axis and Y-axis directions.

[0097] Step S350: Correct the robot arm motion error of the preliminary displacement prediction vector, calculate the centrifugal force offset and Coriolis force interference during the motion process by combining the robot arm kinematic model of the wafer inspection machine, generate the correction coefficient matrix, and perform weighted fusion with the preliminary displacement prediction vector to generate the displacement compensation feature quantity of each particle.

[0098] Robotic arm motion error correction aims to improve the accuracy of displacement compensation by considering the impact of centrifugal force and Coriolis force generated during the robotic arm's movement on particle displacement prediction. The kinematic model of the wafer inspection machine's robotic arm describes its motion laws and mechanical characteristics. This model allows for the calculation of centrifugal force offset and Coriolis force interference during the robotic arm's movement. The correction coefficient matrix, calculated based on the centrifugal force offset and Coriolis force interference, is used to correct the initial displacement prediction vector. The displacement compensation characteristic quantity is a more accurate measure obtained after correction, used to compensate for particle displacement.

[0099] When correcting the motion error of the robotic arm, the following steps are taken: First, based on the kinematic model of the robotic arm of the wafer inspection machine and its motion parameters, the centrifugal force offset and Coriolis force interference during the motion process are calculated. Then, a correction coefficient matrix is ​​generated based on these offsets and interferences. Finally, the correction coefficient matrix is ​​weighted and fused with the preliminary displacement prediction vector. For example, the corresponding elements of the correction coefficient matrix and the preliminary displacement prediction vector are multiplied to obtain the displacement compensation characteristic quantity of each particle.

[0100] In one implementation, step S350 may include the following steps S351 to S355: Step S351: Obtain the set of motion parameters of the robotic arm of the wafer inspection machine. The set of motion parameters of the robotic arm includes the real-time angular velocity value, angular acceleration value of each joint and the current position coordinates of the end effector of the robotic arm.

[0101] The set of motion parameters for a robotic arm is a set of parameters that describe the motion state of the robotic arm. The real-time angular velocity value represents the rotational speed of each joint of the robotic arm at the current moment, the angular acceleration value represents the rate of change of the joint rotational speed, and the current position coordinates of the robotic arm end effector represent the position of the robotic arm end effector in space.

[0102] When acquiring the set of motion parameters for the robotic arm, sensors installed on each joint of the robotic arm, such as angular velocity sensors and angular acceleration sensors, can be used to collect the angular velocity and angular acceleration values ​​of each joint in real time. At the same time, the current position coordinates of the robotic arm's end effector are obtained through position sensors. These parameters are then combined to form the set of motion parameters for the robotic arm.

[0103] Step S352: Based on the kinematic model of the robotic arm, calculate the centrifugal force acceleration vector and Coriolis force acceleration vector of the end effector of the robotic arm during the motion process according to the set of motion parameters of the robotic arm. The centrifugal force acceleration vector is proportional to the square of the joint angular velocity of the robotic arm, and the Coriolis force acceleration vector is proportional to the product of the joint angular velocity and angular acceleration of the robotic arm.

[0104] The kinematic model of a robotic arm is a mathematical model established based on the structure and motion principles of the robotic arm. It can describe the motion relationships and mechanical characteristics of each joint of the robotic arm and can use existing common models. The centrifugal acceleration vector is the acceleration vector caused by the centrifugal force generated by the rotational motion of the robotic arm joints, and is proportional to the square of the angular velocity of the robotic arm joints. The Coriolis acceleration vector is the acceleration vector caused by the Coriolis force generated by the rotational motion of the robotic arm joints and the relative motion of the object in the rotating coordinate system, and is proportional to the product of the angular velocity and angular acceleration of the robotic arm joints. When calculating the centrifugal and Coriolis acceleration vectors, the angular velocity values, angular acceleration values, and the position coordinates of the end effector from the robotic arm motion parameter set are input into the robotic arm kinematic model. Based on the calculation formulas for centrifugal force and Coriolis force in the model, the centrifugal acceleration vector and Coriolis acceleration vector of the robotic arm end effector during its motion are calculated.

[0105] Step S353: Convert the centrifugal force acceleration vector and the Coriolis force acceleration vector into displacement offsets. Combine the motion time parameters of the robotic arm to calculate the centrifugal force offset and Coriolis force interference of the end effector of the robotic arm. The centrifugal force offset is calculated based on the centrifugal force acceleration vector, the square of the motion time parameter, and the dynamic parameters of the robotic arm. The Coriolis force interference is calculated based on the Coriolis force acceleration vector, the square of the motion time parameter, and the dynamic parameters of the robotic arm.

[0106] Displacement offset is the conversion of an acceleration vector into a change in displacement over a certain time interval. The robotic arm motion time parameter refers to the time it takes for the robotic arm to move from its initial position to its current position. Centrifugal force offset is the displacement shift caused by centrifugal acceleration over a certain time interval, calculated using the centrifugal acceleration vector, the square of the motion time parameter, and the robotic arm's dynamic parameters (such as mass and moment of inertia). Coriolis force disturbance is the displacement disturbance caused by Coriolis force acceleration over a certain time interval, calculated using the Coriolis force acceleration vector, the square of the motion time parameter, and the robotic arm's dynamic parameters.

[0107] During the conversion and calculation, based on the kinematic formula: displacement offset = 0.5 * acceleration vector * square of motion time parameter * robot arm dynamic parameter, the centrifugal force acceleration vector and Coriolis force acceleration vector are converted into centrifugal force offset and Coriolis force disturbance, respectively.

[0108] Step S354: Construct an error covariance matrix based on the centrifugal force offset and the Coriolis force interference. Iteratively update the error covariance matrix using the Kalman filter algorithm to generate a correction coefficient matrix containing the X-axis correction coefficient and the Y-axis correction coefficient.

[0109] The error covariance matrix describes the error correlation between centrifugal force offset and Coriolis force disturbance. The Kalman filter algorithm is used for optimal estimation and filtering of the system state. It iteratively updates the error covariance matrix based on the system's dynamic model and observation data, improving estimation accuracy. The correction coefficient matrix, obtained after updating using the Kalman filter algorithm, includes X-axis and Y-axis correction coefficients and is used to correct the initial displacement prediction vector.

[0110] When constructing the error covariance matrix, the covariance between the centrifugal force offset and the Coriolis force disturbance is calculated based on their statistical characteristics. Then, the error covariance matrix is ​​used as the initial input to the Kalman filter algorithm. Combined with observational data of the robotic arm's motion (such as the actual position of the end effector), the error covariance matrix is ​​continuously adjusted through the iterative update process of the Kalman filter algorithm, ultimately generating a correction coefficient matrix containing correction coefficients for the X-axis and Y-axis directions.

[0111] Step S355: Multiply the X-axis direction correction coefficient in the correction coefficient matrix with the X-axis direction displacement component in the preliminary displacement prediction vector, and multiply the Y-axis direction correction coefficient with the Y-axis direction displacement component in the preliminary displacement prediction vector to generate the corrected displacement component. Combine the corrected X-axis direction displacement component and the Y-axis direction displacement component to form the displacement compensation characteristic quantity of each particle.

[0112] By weighted and fused with the correction coefficient matrix and the initial displacement prediction vector, the initial displacement prediction vector can be corrected to obtain a more accurate displacement compensation characteristic quantity. The corrected displacement component is the result of multiplying the correction coefficient with the initial displacement component. Combining the corrected X-axis displacement component and the Y-axis displacement component forms the displacement compensation characteristic quantity of each particle.

[0113] During the calculation, the X-axis correction coefficient in the correction coefficient matrix is ​​multiplied by the X-axis displacement component in the initial displacement prediction vector to obtain the corrected X-axis displacement component. Similarly, the Y-axis correction coefficient is multiplied by the Y-axis displacement component in the initial displacement prediction vector to obtain the corrected Y-axis displacement component. Combining these two corrected displacement components yields the displacement compensation characteristic quantity for each particle.

[0114] Step S400: Call the fine-tuning optical system of the wafer inspection machine to perform optical correction response on the displacement compensation feature quantity, and generate corrected real-time imaging data. The real-time imaging data is a corrected imaging sequence after field distortion adjustment and focus drift compensation.

[0115] The fine-tuning optical system is an optical device used in wafer inspection machines to fine-tune the imaging system, adjusting parameters such as the field of view and focus. Optical correction response refers to the adjustments made by the fine-tuning optical system to the imaging system based on displacement compensation characteristics to compensate for particle positional shifts. Field-of-view distortion adjustment corrects potential distortions in the imaging field of view, making the image more accurately reflect the actual scene. Focus drift compensation compensates for potential focus drift, ensuring image sharpness. The corrected imaging sequence is the imaging data sequence obtained after field-of-view distortion adjustment and focus drift compensation, and is used as real-time imaging data.

[0116] During optical correction response, displacement compensation features are input into the fine-tuning optical system. Based on the displacement components in the X and Y axes of the displacement compensation features, the fine-tuning optical system adjusts the field of view of the imaging system to compensate for particle positional shifts. Simultaneously, based on the displacement compensation features and the height features in the initial particle coordinate set, the focus of the imaging system is adjusted to compensate for focus drift. Finally, a corrected imaging sequence, adjusted for field distortion and compensated for focus drift, is generated as the corrected real-time imaging data.

[0117] As one implementation, step S400 may include the following steps S410~S450: Step S410: Analyze the X-axis displacement component and Y-axis displacement component in the displacement compensation feature quantity, and generate a correction control command for the fine-tuning optical system. The correction control command includes X-axis fine-tuning parameters and Y-axis fine-tuning parameters, and the unit of the fine-tuning parameters matches the stepper motor control accuracy of the fine-tuning optical system.

[0118] Analyzing the displacement compensation characteristic involves extracting the X-axis and Y-axis displacement components to generate correction control commands. These commands are used to control the fine-tuning optical system, where the X-axis and Y-axis fine-tuning parameters represent the required fine-tuning amounts in the X and Y axes, respectively. The units of these fine-tuning parameters must match the control precision of the stepper motor in the fine-tuning optical system to ensure accurate displacement adjustment.

[0119] When analyzing the displacement compensation characteristic quantities, the X-axis and Y-axis displacement components are directly obtained from them. Then, based on the stepper motor control accuracy of the fine-tuning optical system, the displacement components are converted into appropriate fine-tuning parameters. For example, if the stepper motor control accuracy is 0.1 mm and the displacement component is 1 mm, then the corresponding fine-tuning parameter is 10 step units. These fine-tuning parameters are then combined into a correction control command.

[0120] Step S411: Send the correction control command to the piezoelectric drive unit of the fine-tuning optical system to drive the X-axis translation stage and Y-axis translation stage of the fine-tuning optical system to perform micro-displacement adjustment. The displacement adjustment amount of the translation stage is linearly related to the fine-tuning parameters in the correction control command.

[0121] The piezoelectric drive unit is a device in a fine-tuning optical system used to drive the translation stage for displacement adjustment. It utilizes the properties of piezoelectric materials to convert electrical signals into mechanical displacement. The X-axis and Y-axis translation stages are platforms used to adjust the position of the imaging system along the X and Y axes, respectively. The displacement adjustment amount is linearly related to the fine-tuning parameters in the calibration control command; that is, the larger the fine-tuning parameters, the greater the displacement adjustment amount of the translation stage.

[0122] During displacement adjustment, a correction control command is sent to the piezoelectric drive unit of the fine-tuning optical system. Based on the X-axis and Y-axis fine-tuning parameters in the correction control command, the piezoelectric drive unit drives the X-axis and Y-axis translation stages to perform micro-displacement adjustments, respectively. For example, if the X-axis fine-tuning parameter is 5 steps, the X-axis translation stage will move a certain distance accordingly based on a linear correspondence.

[0123] Step S412: Call the laser interferometer of the fine-tuning optical system to perform real-time feedback detection of the actual displacement of the translation stage, and generate a displacement feedback signal. The displacement feedback signal includes the actual displacement value in the X-axis direction and the actual displacement value in the Y-axis direction.

[0124] A laser interferometer is a high-precision measuring device that uses the interference principle of lasers to measure the displacement of an object. Real-time feedback detection refers to continuously measuring the actual displacement of the translation stage during displacement adjustment and feeding the measurement results back. The displacement feedback signal contains the actual displacement values ​​of the translation stage in the X and Y axis directions, which allows us to understand the actual adjustment status of the translation stage.

[0125] During real-time feedback detection, the laser interferometer of the fine-tuning optical system is activated to measure the displacement of the translation stages in the X and Y axes. The laser interferometer converts the measured actual displacement values ​​in the X and Y axes into electrical signals, generating a displacement feedback signal.

[0126] Step S413: Compare the deviation between the displacement feedback signal and the fine-tuning parameters in the correction control command, calculate the displacement adjustment deviation, and when the displacement adjustment deviation exceeds the preset deviation threshold, generate a secondary correction control command to drive the piezoelectric drive unit to perform compensation adjustment until the displacement adjustment deviation is less than the preset deviation threshold.

[0127] Deviation comparison involves comparing the actual X-axis and Y-axis displacement values ​​contained in the displacement feedback signal with the X-axis and Y-axis fine-tuning parameters in the correction control command, respectively, to determine whether the actual displacement of the translation stage has reached the expected adjustment amount. The displacement adjustment deviation is the difference between the actual displacement value and the fine-tuning parameters, reflecting the degree of error in the translation stage's displacement adjustment. The preset deviation threshold is a pre-set maximum allowable deviation value. If the displacement adjustment deviation exceeds this threshold, it indicates that the translation stage's displacement adjustment has not achieved the expected effect and further adjustment is required. The secondary correction control command is a new control command generated based on the displacement adjustment deviation, used to drive the piezoelectric drive unit to perform additional compensation adjustments on the translation stage.

[0128] When comparing deviations, the displacement adjustment deviations in the X-axis and Y-axis directions are calculated separately. This involves calculating the difference between the actual displacement value and the fine-tuning parameter in the X-axis direction, and the difference between the actual displacement value and the fine-tuning parameter in the Y-axis direction. If the calculated displacement adjustment deviation exceeds a preset deviation threshold in either the X-axis or Y-axis direction, a secondary correction control command is generated based on the magnitude and direction of the deviation. For example, if the displacement adjustment deviation in the X-axis direction is positive and exceeds the threshold, it indicates that the actual displacement of the translation stage in the X-axis direction exceeds expectations. In this case, the secondary correction control command instructs the piezoelectric drive unit to make a small, reverse displacement adjustment of the translation stage in the X-axis direction. By continuously comparing deviations and generating secondary correction control commands, the piezoelectric drive unit is driven to perform compensation adjustments until the displacement adjustment deviation is less than the preset deviation threshold in both the X-axis and Y-axis directions. At this point, the displacement adjustment of the translation stage is considered to have achieved the expected effect.

[0129] Step S450: Control the imaging lens group of the fine-tuning optical system to adjust the focus, calculate the focus compensation amount based on the height feature quantity in the initial particle coordinate set and the displacement component in the displacement compensation feature quantity, drive the focusing motor of the lens group to move axially, generate the imaging optical path after focus adjustment, and collect the corrected imaging sequence after field distortion adjustment and focus drift compensation as real-time imaging data.

[0130] The imaging lens assembly is a key component in the fine-tuning optical system used for imaging. Focus adjustment is crucial for the imaging system to clearly capture particles on the wafer surface. The focus compensation amount is calculated based on the height characteristic value in the initial particle coordinate set and the displacement component in the displacement compensation characteristic value, used to compensate for focus drift caused by changes in particle height and positional shift. The focusing motor is used to drive the imaging lens assembly axially. By controlling the movement of the focusing motor, the focus position of the imaging lens assembly can be adjusted. The imaging optical path is the path formed by light rays passing through the imaging lens assembly. The focused imaging optical path ensures clear imaging at the new focus position. The corrected imaging sequence is the sequence of imaging data acquired after field-of-view distortion adjustment and focus drift compensation. This sequence is used as real-time imaging data for accurate analysis of the wafer particle positions.

[0131] When adjusting the focus, the height characteristics in the initial particle coordinate set are first analyzed to understand the overall height distribution of particles on the wafer surface. Simultaneously, the displacement component in the displacement compensation characteristics is considered, as particle position shifts may cause changes in the imaging optical path, thus affecting the focus position. The focus compensation amount is calculated by combining the height characteristics and the displacement component. For example, a mathematical model can be established based on optical principles, considering the impact of particle height changes and position shifts on the focus. By inputting the height characteristics and displacement component, the focus compensation amount is output. Then, the focus compensation amount is converted into control parameters for the focusing motor, driving the focusing motor to move axially, moving the imaging lens assembly to the new focus position, and generating the focused-adjusted imaging optical path. Finally, the imaging unit of the fine-tuning optical system is controlled to acquire images at certain time intervals, obtaining a corrected imaging sequence after field-of-view distortion adjustment and focus drift compensation, which is then used as real-time imaging data.

[0132] In one implementation, in step S450, the imaging lens group of the fine-tuning optical system is controlled to adjust the focus. The focus compensation amount is calculated based on the height feature quantity in the initial particle coordinate set and the displacement component in the displacement compensation feature quantity. The focusing motor of the lens group is driven to move axially to generate the imaging optical path after focus adjustment. Specifically, it may include the following steps S451~S455: Step S451: Statistical analysis is performed on the height feature quantity in the initial particle coordinate set, and the distribution morphology feature parameters of all particle height feature quantities are calculated. The distribution morphology feature parameters are used to characterize the overall height distribution trend of particles on the wafer surface.

[0133] Statistical analysis is the process of processing and analyzing height characteristics in an initial set of particle coordinates. By calculating distribution morphology parameters, a more comprehensive understanding of the height distribution of particles on the wafer surface can be obtained. Distribution morphology parameters include central tendency parameters (such as mean, median, etc.) and dispersion parameters (such as standard deviation, variance, etc.). These parameters can describe the distribution pattern of particle height from different perspectives. For example, the mean can reflect the average level of particle height, while the standard deviation can reflect the degree of dispersion of particle height relative to the mean.

[0134] In the statistical analysis, all height features in the initial particle coordinate set are first extracted to form a dataset. Then, statistical methods are used to calculate the distribution morphology parameters of this dataset. For the central tendency parameter, when calculating the mean, all height features are summed and then divided by the total number of particles; when calculating the median, the height features are arranged in ascending order, and the middle value is taken (if the total number of particles is even, the average of the two middle values ​​is taken). For the dispersion parameter, when calculating the standard deviation, the square of the difference between each height feature and the mean is calculated, these squares are summed and divided by the total number of particles, and then the square root is taken; when calculating the variance, the square of the standard deviation is used directly. These calculated distribution morphology parameters effectively characterize the overall height distribution trend of particles on the wafer surface.

[0135] Step S452: Calculate the initial focal position parameter based on the central trend parameter in the distribution morphology characteristic parameters, set the initial focal point of the imaging lens group to the axial position corresponding to the central trend parameter of the height characteristic quantity, and generate the initial imaging optical path parameters.

[0136] The center tendency parameter reflects the average or typical value of particle height. Calculating the initial focal position parameter based on these parameters allows the initial focus of the imaging lens assembly to be roughly aligned with the center height of the particles on the wafer surface, improving image sharpness. The initial focal position parameter refers to the axial position that the imaging lens assembly should be in the initial state, corresponding to the center tendency parameter of the height characteristic. The initial imaging optical path parameters are a set of parameters describing the imaging optical path in the initial state, including the lens focal length, object distance, and image distance. These parameters determine the propagation path of light through the imaging lens assembly and the imaging effect.

[0137] When calculating the initial focal position parameters, the initial focal position is obtained through a specific formula based on the central trend parameter (such as the mean or median), combined with the optical characteristics and geometric relationships of the imaging system. For example, the object distance corresponding to the central trend parameter can be calculated using the lens imaging formula 1 / f = 1 / u + 1 / v (where f is the focal length, u is the object distance, and v is the image distance), and the correspondence between particle height and object distance. This allows the axial position of the imaging lens assembly to be determined as the initial focal position parameter. Then, the imaging lens assembly is adjusted to this axial position, and other relevant optical parameters are determined based on the imaging system design and current settings to generate the initial imaging optical path parameters. In this way, the imaging lens assembly is in a preliminary focal position, capable of clearly capturing particles on the wafer surface to a certain extent.

[0138] Step S453: Calculate the combined plane displacement based on the X-axis displacement component and the Y-axis displacement component in the displacement compensation characteristic quantity, and calculate the square root of the sum of the squares of the X-axis and Y-axis displacement components as the combined plane displacement.

[0139] The overall planar displacement is an indicator used to measure the overall displacement of a particle in a plane, taking into account the displacement components along the X and Y axes. By calculating the square root of the sum of the squares of the X and Y axis displacement components, a numerical value reflecting the actual displacement of the particle in the plane can be obtained. This calculation method is based on the Pythagorean theorem, treating the displacements along the X and Y axes as the two legs of a right triangle, and the overall planar displacement is the length of the hypotenuse of this right triangle.

[0140] When calculating the total planar displacement, the X-axis and Y-axis displacement components from the displacement compensation characteristic quantity are directly obtained. Assuming the X-axis displacement component is x and the Y-axis displacement component is y, the total planar displacement d can be calculated using the formula d = √(x / y). 2 +y 2 The calculation yields the result. For example, if the displacement component along the X-axis is 3 units and the displacement component along the Y-axis is 4 units, then the total planar displacement d = √(3 2 +4 2=5 units. This comprehensive planar displacement can accurately reflect the overall displacement of the particles on the plane.

[0141] Step S454: Based on the comprehensive plane displacement, query the preset displacement-focus offset mapping table to obtain the focus offset coefficient caused by the plane displacement. The displacement-focus offset mapping table is constructed through the previous calibration experiment and contains the focus offset coefficients corresponding to different comprehensive plane displacements.

[0142] The displacement-focal offset mapping table is a pre-built table that records the mapping relationship between different combined planar displacements and their corresponding focal offset coefficients. The focal offset coefficient refers to the degree of shift in the focal position caused by the particle's displacement on the plane; it is a coefficient related to the combined planar displacement. The preliminary calibration experiments were conducted to determine the displacement-focal offset mapping table. In these experiments, different planar displacement conditions were tested, and the corresponding focal offsets were measured, thereby establishing the correspondence between the combined planar displacement and the focal offset coefficients.

[0143] When querying the displacement-focus offset mapping table, the calculated total planar displacement is used as the search keyword. The table is then searched for the record closest to this total planar displacement, and the corresponding focus offset coefficient is obtained. For example, if the calculated total planar displacement is 5 units, and the table finds a focus offset coefficient of 0.2 corresponding to 5 units, then 0.2 is used as the focus offset coefficient caused by the current planar displacement. This focus offset coefficient will be used to calculate the focus compensation amount to further adjust the focus position of the imaging lens assembly.

[0144] Step S455: Perform optical coupling calculation on the focal offset coefficient and the dispersion parameter of the height feature quantity. Based on the influence of the optical path tilt caused by the plane displacement and the dispersion of the height distribution on the focal depth, generate the focal compensation quantity. The optical coupling calculation is realized through a pre-constructed optical transfer function, which characterizes the joint influence weight of the plane displacement and the height dispersion on the focal position.

[0145] Optical coupling calculation is a process that comprehensively considers the influence of optical path tilt caused by planar displacement and height distribution dispersion on focal depth, combining the focal shift coefficient and the dispersion parameter of height characteristics. Optical path tilt refers to the degree of tilt of the light propagation path caused by the displacement of particles on the plane; height distribution dispersion refers to the dispersion of particle height on the wafer surface; and focal depth refers to the axial range within which a clear image can be formed in the imaging system. The focal compensation amount, obtained through optical coupling calculation, is used to compensate for focal drift caused by changes in planar displacement and height distribution, enabling the imaging system to maintain clear imaging under new conditions. The pre-constructed optical transfer function is a mathematical function describing the joint influence of planar displacement and height dispersion on the focal position, containing weighting coefficients for different factors. This function can be used to accurately calculate the focal compensation amount.

[0146] During optical coupling calculations, the focus offset coefficient and the dispersion parameters (such as standard deviation) of height features are input into a pre-constructed optical transfer function (OPF). The OPF processes and calculates the input parameters based on its internal mathematical model and weighting coefficients. For example, the OPF might be a linear combination: focus compensation = a * focus offset coefficient + b * dispersion parameter of height features (where a and b are predefined weighting coefficients). In this way, the influence of planar displacement and height distribution on the focus position is comprehensively considered to generate the final focus compensation. Then, based on the focus compensation, the focusing motor is driven to move axially, adjusting the focus position of the imaging lens group and generating the focused imaging optical path.

[0147] As one implementation, in step S450, the corrected imaging sequence after field distortion adjustment and focus drift compensation is acquired as real-time imaging data, which may specifically include the following steps S456~S4510: Step S456: Control the first view imaging unit and the second view imaging unit of the fine-tuning optical system to perform synchronous exposure acquisition, and generate the first corrected imaging sequence and the second corrected imaging sequence after field distortion adjustment. The timestamps of the first corrected imaging sequence and the second corrected imaging sequence are kept synchronized.

[0148] Synchronous exposure acquisition refers to controlling the first-view imaging unit and the second-view imaging unit of the fine-tuning optical system to perform exposure operations simultaneously, ensuring that the acquired images are consistent in time. The first-view imaging unit and the second-view imaging unit image the wafer surface from different angles. After the field-of-view distortion adjustment and focus drift compensation steps in the previous steps, the acquired images can more accurately reflect the actual situation of the wafer surface particles. The first and second corrected imaging sequences are the imaging data sequences acquired after adjustment. Timestamps are used to record the acquisition time of each imaging frame. Keeping the timestamps synchronized facilitates the matching and analysis of the images in the two sequences.

[0149] During synchronized exposure acquisition, the control module of the fine-tuning optical system sends synchronization trigger signals to the first-view imaging unit and the second-view imaging unit, causing them to begin exposure simultaneously. During exposure, the imaging units continuously acquire images at a certain frame rate, forming a first corrected imaging sequence and a second corrected imaging sequence. Simultaneously, an accurate timestamp is added to each imaging frame to ensure that the timestamps of the two sequences remain synchronized. For example, a high-precision clock source can be used to provide a time signal to the imaging units, ensuring the accuracy of the timestamps.

[0150] Step S457: Evaluate the image sharpness of the first and second corrected imaging sequences. Calculate the sharpness evaluation value of each imaging frame using the Laplacian operator and select valid imaging frames with sharpness evaluation values ​​higher than a preset sharpness threshold.

[0151] Image sharpness evaluation assesses the sharpness of each frame in the first and second corrected imaging sequences to select high-quality images for analysis. The Laplacian operator is a commonly used operator for detecting image edges and sharpness, reflecting local changes in the image by calculating the second derivative of pixels. The sharpness evaluation value is a numerical value obtained after processing the imaging frame using the Laplacian operator; a higher value indicates a sharper image. A preset sharpness threshold is a pre-defined standard value; only imaging frames with sharpness evaluation values ​​higher than this threshold are considered valid.

[0152] During image sharpness evaluation, the Laplacian operator is applied to each imaging frame in the first and second corrected imaging sequences. Specifically, the Laplacian operator is convolved with the imaging frame to obtain a new image. The sum of the absolute values ​​of all pixels in this new image is then calculated and used as the sharpness evaluation value for that imaging frame. The sharpness evaluation value of each imaging frame is compared to a preset sharpness threshold. If the evaluation value is higher than the threshold, the imaging frame is considered a valid imaging frame and is selected. For example, if the preset sharpness threshold is 100, and the sharpness evaluation value of an imaging frame is 120, then that imaging frame is selected as a valid imaging frame.

[0153] Step S458: Perform bad pixel repair processing on the effective imaging frame. Use the median filtering algorithm to filter isolated noise points in the imaging frame, preserve the detailed features of the particle area, and generate a noise-reduced corrected imaging frame.

[0154] Defect pixel repair aims to remove isolated noise pixels that may exist in the effective imaging frame, thereby improving image quality. Isolated noise pixels refer to single pixels in an image that differ significantly from surrounding pixels. These may be caused by noise from the imaging device, sensor malfunction, or other reasons, and can affect image sharpness and analysis results. Median filtering is a commonly used non-linear filtering algorithm. By sorting the pixel values ​​within a window and taking the median value as the new value for the center pixel, it effectively removes isolated noise pixels while preserving the image's edge and detail features. The denoised and corrected imaging frame is the image obtained after median filtering. While removing noise, it retains the detailed information of the wafer surface grain region, facilitating analysis and processing.

[0155] In the process of bad pixel repair, a mean-mode filtering algorithm is applied to each valid imaging frame. First, a suitable filtering window size is selected, such as a 3x3 or 5x5 window. Then, the filtering window is moved pixel by pixel across the imaging frame, and the pixel values ​​within each window are sorted, with the median value taken as the new value for the center pixel. In this way, each pixel in the imaging frame is processed to remove isolated noise points. During processing, care is taken to preserve the detailed features of the grainy areas to avoid over-filtering that could lead to the loss of grain information. For example, for pixels at grain edges, the size of the filtering window is appropriately adjusted or an adaptive filtering method is used to ensure that the edge information of the grains is preserved. Finally, a noise-reduced corrected imaging frame is generated.

[0156] Step S459: Perform contrast enhancement processing on the noise-reduced corrected imaging frame. Adjust the gray-scale dynamic range of the imaging frame through an adaptive histogram equalization algorithm to enhance the gray-scale contrast between the grain area and the background area, and generate an enhanced corrected imaging frame.

[0157] Contrast enhancement is performed to improve the grayscale difference between grain and background regions in the denoised corrected imaging frame, making the grains more clearly distinguishable. Adaptive histogram equalization is an algorithm that adaptively adjusts the histogram based on local image features, enhancing image contrast while avoiding over-enhancement or detail loss that can occur with traditional histogram equalization. Grayscale dynamic range refers to the range of grayscale values ​​of pixels in an image; adjusting the grayscale dynamic range makes the grayscale values ​​of grain and background regions more clearly distinguishable. The enhanced corrected imaging frame is the image obtained after contrast enhancement, possessing higher grayscale contrast and more clearly displaying the grain information on the wafer surface.

[0158] During contrast enhancement, an adaptive histogram equalization algorithm is applied to each denoised and corrected imaging frame. First, the imaging frame is divided into multiple small local regions. Then, histogram equalization is performed on each local region separately, calculating its histogram and adjusting the grayscale values ​​of the pixels within that region based on the histogram. During this adjustment process, the equalization parameters are adaptively determined based on the statistical information of the local region to ensure that image detail is preserved while enhancing contrast. Finally, the processed local regions are merged into a complete image to obtain the enhanced and corrected imaging frame. This method enhances the grayscale contrast between the grainy and background regions, making the grains more prominent.

[0159] Step S4510: Reassemble the enhanced and corrected imaging frames according to the timestamp order to generate a corrected imaging sequence after field distortion adjustment and focus drift compensation, and determine the corrected imaging sequence as real-time imaging data.

[0160] Sequence reassembly involves rearranging the contrast-enhanced and corrected imaging frames according to their timestamps to form an ordered imaging sequence. The timestamps record the acquisition time of each imaging frame; reassembling according to timestamp order ensures the temporal continuity of the corrected imaging sequence, consistent with the actual imaging process. The corrected imaging sequence, after field-of-view distortion adjustment and focus drift compensation, is the final imaging data sequence. It integrates the results of multiple previous steps, including field-of-view distortion adjustment, focus drift compensation, image sharpness filtering, bad pixel repair, and contrast enhancement, accurately reflecting the actual situation of the wafer surface particles. This corrected imaging sequence is designated as real-time imaging data for coordinate registration and final position determination of the wafer particles.

[0161] During sequence recombination, the timestamp of each enhanced and corrected imaging frame is first extracted. Then, the enhanced and corrected imaging frames are sorted according to their timestamps, arranged in ascending chronological order. Finally, the sorted enhanced and corrected imaging frames are combined into a new imaging sequence, which is the corrected imaging sequence after field-of-view distortion adjustment and focus drift compensation. This corrected imaging sequence is used as real-time imaging data and input into the coordinate registration step to provide the basic data for accurately determining the final position of the wafer grains.

[0162] Step S500: Perform coordinate registration on the real-time imaging data to generate the final position coordinate set of the wafer particles. The final position coordinate set is used to indicate the actual distribution position of each particle on the wafer surface.

[0163] Coordinate registration is the process of aligning and matching images from different viewpoints or at different times in real-time imaging data, ensuring that the coordinates of the same object in different images correspond. Coordinate registration eliminates positional discrepancies between images, accurately determining the position of wafer particles in a unified coordinate system. The final set of position coordinates is the coordinate data set obtained after coordinate registration, containing the actual distribution positions of each particle on the wafer surface.

[0164] When performing coordinate registration, the first step is to select a suitable registration method and feature points. Image features in real-time imaging data, such as edges and corners, can be used as the basis for registration. A feature point matching algorithm is used to find corresponding feature point pairs in images from different viewpoints. Then, a coordinate transformation model is established based on these feature point pairs. By solving for the model's parameters, the coordinates in different images are transformed into a unified coordinate system. Finally, the transformed coordinates are combined into the final set of position coordinates. For example, an affine transformation model can be used for coordinate registration. By solving for the rotation, translation, and scaling parameters of the affine transformation using the least squares method, the particle coordinates in the real-time imaging data are transformed into the global coordinate system of the wafer surface, generating the final set of position coordinates.

[0165] As one implementation, step S500 may include the following steps S510~S550: Step S510: Extract the first corrected imaging sequence and the second corrected imaging sequence from the real-time imaging data, perform sub-pixel level feature point matching on the two corrected imaging sequences, and generate a set of registered feature point pairs. The set of registered feature point pairs contains feature point coordinate pairs with spatial correspondence.

[0166] Subpixel-level feature point matching performs feature point matching at a finer level than pixels, improving matching accuracy. The first and second corrected imaging sequences are imaging sequences acquired from different perspectives in real-time imaging data. By performing subpixel-level feature point matching on them, the precise positional relationship of corresponding particles in the two sequences can be found. The registration feature point pair set is a collection of feature point coordinate pairs with spatial correspondences; these coordinate pairs are used to establish a coordinate transformation model.

[0167] In subpixel-level feature point matching, feature points are first extracted from the first and second corrected imaging sequences. Feature point detection algorithms such as SIFT and SURF can be used to find feature points in both sequences. Then, these feature points are located at the subpixel level by fitting the grayscale distribution of the local image to pinpoint their location to the subpixel level. Next, feature descriptors, such as SIFT descriptors, are used to describe the feature points. By comparing the similarity of the feature descriptors, corresponding feature point pairs in the two sequences are found. These spatially corresponding feature point coordinate pairs are combined into a registration feature point pair set. For example, if a feature point with coordinates (x1, y1) is found in the first corrected imaging sequence, and its corresponding feature point with coordinates (x2, y2) is found in the second corrected imaging sequence, these two coordinates are combined into a coordinate pair (x1, y1, x2, y2) and added to the registration feature point pair set.

[0168] Step S520: Construct a coordinate transformation model based on the set of registered feature points, and solve the transformation parameters of the coordinate transformation model by the least squares method. The coordinate transformation model is an affine transformation model, which includes rotation parameters, translation parameters and scaling parameters.

[0169] Coordinate transformation models are used to convert the coordinates of one image to the coordinate system of another image, thereby achieving image registration. Affine transformation models are commonly used coordinate transformation models that can describe transformations such as rotation, translation, and scaling of images. Rotation parameters represent the angle of rotation of the image around a point, translation parameters represent the amount of horizontal and vertical movement of the image, and scaling parameters represent the scaling ratio of the image in the horizontal and vertical directions. The least squares method is an optimization method used to solve systems of linear equations, determining the optimal parameters of the coordinate transformation model by minimizing the sum of squared errors.

[0170] When constructing the coordinate transformation model, the mathematical equations for the affine transformation are established based on the coordinate pairs in the registration feature point set. An affine transformation can be represented as a system of linear equations, containing rotation, translation, and scaling parameters. Substituting the coordinate pairs from the registration feature point set into this system of equations yields a set of linear equations. These equations are then solved using the least squares method. By minimizing the sum of squared errors between the actual coordinates and the transformed coordinates, the rotation, translation, and scaling parameters of the affine transformation model are determined.

[0171] Step S530: Map the particle region coordinates in the first corrected imaging sequence to the imaging coordinate system of the second corrected imaging sequence through a coordinate transformation model to generate a set of particle registration coordinates after coordinate mapping.

[0172] Mapping the particle region coordinates in the first corrected imaging sequence using a coordinate transformation model aims to unify the particle coordinates from different viewpoints into a single coordinate system, facilitating analysis and processing. The mapped particle registration coordinate set is obtained by performing an affine transformation on the particle region coordinates in the first corrected imaging sequence, and these coordinates correspond to the imaging coordinate system of the second corrected imaging sequence.

[0173] During coordinate mapping, for each particle region coordinate (x, y) in the first corrected imaging sequence, it is substituted into the already solved affine transformation model. Based on the mathematical equations of the affine transformation, the corresponding coordinates (x', y') in the imaging coordinate system of the second corrected imaging sequence are calculated. After mapping all particle region coordinates, the set of particle registration coordinates after coordinate mapping is obtained. For example, given the parameter a of the affine transformation model... 11 ,a 12 ,a 21 ,a 22 ,t x ,t y For the particle region coordinates (x, y) in the first corrected imaging sequence, x'=a is calculated. 11 x+a 12 y+t x and y'=a 21 x+a 22 y+t y The corresponding registration coordinates (x', y') are obtained and added to the particle registration coordinate set.

[0174] Step S540: Perform consistency verification on the particle registration coordinate set, calculate the area deviation rate and centroid deviation distance of the particle region before and after mapping, remove abnormal registration coordinates with area deviation rate exceeding the preset area threshold, and remove abnormal registration coordinates with centroid deviation distance exceeding the preset distance threshold.

[0175] Consistency verification is performed to ensure the accuracy and reliability of coordinates in the particle registration coordinate set and to eliminate potential registration errors. Area deviation rate refers to the rate of change in the area of ​​the particle region before and after mapping, reflecting the change in the particle's area during registration. Centroid deviation distance refers to the distance between the centroid positions of the particle region before and after mapping, reflecting the degree of positional shift of the particle during registration. Preset area thresholds and preset distance thresholds are pre-defined standard values ​​used to determine whether the registered coordinates are abnormal. If the area deviation rate exceeds the preset area threshold or the centroid deviation distance exceeds the preset distance threshold, the registered coordinates are considered abnormal and need to be removed.

[0176] During consistency verification, for each registration coordinate in the particle registration coordinate set, the area and centroid coordinates of the particle region in the first corrected imaging sequence before mapping, and the area and centroid coordinates of the corresponding particle region in the second corrected imaging sequence after mapping are first calculated. Then, the area deviation rate is calculated, i.e., area deviation rate = |area after mapping - area before mapping| / area before mapping. Simultaneously, the centroid deviation distance is calculated, i.e., centroid deviation distance = √((centroid x-coordinate after mapping - centroid x-coordinate before mapping)). 2 +(y-coordinate of the centroid after mapping - y-coordinate of the centroid before mapping) 2 The area deviation rate and centroid deviation distance are compared with preset area thresholds and preset distance thresholds, respectively. If the area deviation rate exceeds the preset area threshold or the centroid deviation distance exceeds the preset distance threshold, the registered coordinate is removed from the particle registration coordinate set. For example, if the preset area threshold is 0.2 and the area deviation rate corresponding to a certain registration coordinate is 0.3, then the registration coordinate is considered abnormal and is removed.

[0177] Step S550: Perform coordinate fusion processing on the particle registration coordinate set that has passed the consistency check, fuse the particle coordinates in the first corrected imaging sequence and the second corrected imaging sequence to generate the final position coordinate set of the wafer particles. The final position coordinate set contains the three-dimensional coordinate feature of each particle in the global coordinate system on the wafer surface.

[0178] Coordinate fusion processing combines the particle coordinates from the first and second corrected imaging sequences within the particle registration coordinate set that has passed consistency verification to obtain more accurate particle position information. The global coordinate system is a unified coordinate system used to describe the actual distribution positions of all particles on the wafer surface. The three-dimensional coordinate features include the x, y coordinates (planar position) and z coordinates (height information) of the particles in the global coordinate system, which can completely represent the actual positions of the particles on the wafer surface.

[0179] During coordinate fusion, methods such as weighted averaging can be used to fuse the particle coordinates in the first and second corrected imaging sequences. For example, for each particle, different weights can be assigned based on the reliability or accuracy of the coordinates in the two sequences, and then the weighted average is calculated as the final coordinates of the particle in the global coordinate system. Simultaneously, the previously calculated particle height information is used as the z-coordinate, which, together with the planar coordinates, forms a three-dimensional coordinate feature. The three-dimensional coordinate features of all particles are then combined to form the final set of position coordinates for the wafer particles.

[0180] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0181] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used for sending and receiving data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0182] In one embodiment, the processor 101 executes the vision-based wafer particle position offset detection and compensation method provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A method for detecting and compensating wafer particle position offset based on visual inspection, characterized in that, The method includes: performing dual-view synchronous imaging on the wafer surface to generate an initial imaging data set containing the surface microstructure, the initial imaging data set consisting of a first imaging sequence output by a first-view imaging unit and a second imaging sequence output by a second-view imaging unit, wherein the first and second imaging sequences have the same timestamp and the imaging fields of view have a preset overlapping area; performing parallax matching and feature segmentation on the initial imaging data set to generate three-dimensional height information of wafer surface particles and an initial particle coordinate set, the initial particle coordinate set containing planar coordinate features of each particle in the imaging coordinate system and the corresponding height features; performing particle displacement prediction and correction based on the initial particle coordinate set and the wafer surface charge distribution features to generate displacement compensation features for each particle; calling the fine-tuning optical system of the wafer inspection machine to perform optical correction response on the displacement compensation features to generate corrected real-time imaging data; and performing coordinate registration on the real-time imaging data to generate the final position coordinate set of the wafer particles.

2. The method according to claim 1, characterized in that, The step of performing disparity matching and feature segmentation on the initial imaging data set to generate three-dimensional height information and an initial particle coordinate set for wafer surface particles includes: performing imaging consistency calibration processing on the first and second imaging sequences in the initial imaging data set; adjusting the imaging distortion parameters of the two imaging sequences based on the feature point correspondence of a preset overlapping region; generating a first calibrated imaging sequence and a second calibrated imaging sequence after distortion calibration; calculating the gray-level co-occurrence matrix on the first and second calibrated imaging sequences; and extracting the texture feature distribution map of the two calibrated imaging sequences in the overlapping region, wherein the texture feature distribution map contains gray-level change frequency parameters of different directional channels. The process involves: performing cross-view feature point matching based on the texture feature distribution map; filtering matching feature point pairs that meet the disparity geometry conditions to generate an initial disparity candidate point set; removing outliers from the initial disparity candidate point set; iteratively filtering interior feature point pairs; calculating the disparity value of each interior feature point pair; generating a dense disparity map; combining the dense disparity map with preset dual-view imaging system baseline parameters; converting and generating the three-dimensional height information of wafer surface particles using the triangulation principle; performing region growth segmentation on the first calibration imaging sequence; extracting the boundary contour features of each particle region; and generating an initial particle coordinate set containing planar coordinate features and height features.

3. The method according to claim 2, characterized in that, The cross-view feature point matching process based on the texture feature distribution map, which filters matching feature point pairs that satisfy disparity geometry conditions and generates an initial disparity candidate point set, includes: performing multi-scale feature extraction on the texture feature distribution map, constructing a feature pyramid structure containing multiple scale levels, extracting scale-invariant feature points at different scale levels, and generating a first feature point set and a second feature point set containing scale parameters and orientation parameters. The first feature point set corresponds to a first calibration imaging sequence, and the second feature point set corresponds to a second calibration imaging sequence. The process also involves calculating feature descriptors for the first and second feature point sets, and then using a dynamic weighted similarity fusion algorithm to perform similarity analysis on the feature descriptors of the first and second feature point sets. A similarity metric is performed to generate an initial feature matching matrix. An epipolar geometric model is constructed based on the intrinsic and extrinsic parameter matrices of the dual-view imaging system. The epipolar trajectory corresponding to each first feature point is calculated according to the epipolar geometric model. Candidate matching point pairs located within a preset geometric range of the epipolar trajectory in the initial feature matching matrix are selected. The candidate matching point pairs are subjected to bidirectional consistency verification, and mutually inverse matching point pairs that simultaneously satisfy the optimal matching from the first feature point to the second feature point and from the second feature point to the first feature point are retained. The scale and orientation parameters of the mutually inverse matching point pairs are fused to generate an initial disparity candidate point set with geometric constraints and scale consistency. Each element in the initial disparity candidate point set contains the coordinate information of the matching point pair and the corresponding disparity estimate.

4. The method according to claim 3, characterized in that, The process of performing multi-scale feature extraction on the texture feature distribution map, constructing a feature pyramid structure with multiple scale levels, and extracting scale-invariant feature points at different scale levels to generate a first feature point set and a second feature point set containing scale and orientation parameters, includes: performing multi-scale adaptive smoothing filtering on the texture feature distribution map; performing convolution operations on the texture feature distribution map using Gaussian kernels with different parameters to generate a smooth feature map set with multiple scale levels, where the Gaussian kernel parameters of adjacent scale levels are in a preset ratio; constructing a multi-scale difference feature pyramid in the smooth feature map set by difference operations on adjacent scale maps, where each difference feature pyramid level contains multiple sub-levels, and the map size of each sub-level is a preset ratio of the previous sub-level; and processing the multi-scale... Local extremum detection is performed on each sub-layer of the differential feature pyramid. The feature values ​​of the current detection point are compared with those of its neighbors in the same sub-layer and its upper and lower neighbors in adjacent sub-layers to filter out potential feature points that meet the feature value extremum conditions. Spatial coordinate optimization is performed on the potential feature points. The spatial coordinates and scale parameters of the feature points are calculated by interpolating the feature values ​​of the neighbors. Feature points with feature response intensity lower than a preset feature response threshold and edge response points located at the image boundary are removed. The orientation parameters are calculated based on the neighborhood gradient information of the feature points. The gradient magnitude distribution in different directions is statistically analyzed in the neighborhood of the feature points. A direction histogram is generated by Gaussian weighted smoothing. The direction corresponding to the main peak in the direction histogram is used as the orientation parameter of the feature point to generate a first feature point set and a second feature point set containing scale parameters and orientation parameters.

5. The method according to claim 2, characterized in that, The process of performing region growing and segmentation on the first calibration imaging sequence, extracting the boundary contour features of each particle region, and generating an initial particle coordinate set containing planar coordinate features and height features includes: performing adaptive threshold segmentation on the first calibration imaging sequence, calculating a dynamic segmentation threshold based on the local gray-level mean and standard deviation, and generating a binary imaging mask in which particle regions and background regions have different gray-level distributions; in the binary imaging mask, performing connected region marking using a multi-neighborhood region growing algorithm, using gray-level values ​​as seed point selection criteria, merging adjacent pixels that meet the gray-level similarity threshold into the same connected region, and generating a preliminary particle region marking map; and combining the dense disparity map. The height feature value is used to filter the preliminary particle region marking map, eliminating noise regions with a height feature value lower than a preset height threshold and tiny regions with an area smaller than a preset area threshold, retaining effective particle regions; the effective particle regions are processed by boundary contour extraction to obtain the closed contour curve of each particle region, and the minimum bounding rectangle parameter and centroid coordinates of the contour curve are calculated as planar coordinate feature values; the centroid coordinates of the effective particle regions are associated with the height feature values ​​of the corresponding positions in the dense parallax map to generate an initial particle coordinate set containing planar coordinate feature values ​​and height feature values, and the planar coordinate feature values ​​and height feature values ​​of each particle in the initial particle coordinate set are associated through the same particle identifier.

6. The method according to claim 1, characterized in that, The step of predicting and correcting particle displacement based on the initial particle coordinate set and wafer surface charge distribution characteristics to generate displacement compensation feature quantities for each particle includes: calling the wafer charge detection unit to collect charge distribution characteristics of the wafer surface, wherein the charge distribution characteristics include charge density distribution maps and charge polarity distribution matrices of each region on the wafer surface, and the spatial resolution of the charge density distribution map is consistent with the resolution of the planar coordinate feature quantities of the initial particle coordinate set; performing coordinate transformation processing on the planar coordinate feature quantities of each particle in the initial particle coordinate set, converting the pixel unit planar coordinate feature quantities in the imaging coordinate system into physical length units based on the calibration parameters of the dual-view imaging system, and then mapping them to the charge detection coordinate system of the wafer charge detection unit to generate a particle mapping coordinate set in the charge coordinate system; based on the particle mapping... The system queries the charge density distribution map and charge polarity distribution matrix of the coordinate set, extracts the charge density feature value and charge polarity feature value of the region where each particle is located, and generates a particle charge correlation feature set. Each element in the particle charge correlation feature set corresponds one-to-one with a particle in the initial particle coordinate set. The system calls a pre-trained particle displacement prediction model to perform joint prediction processing on the height feature quantity in the initial particle coordinate set and the particle charge correlation feature set to generate a preliminary displacement prediction vector for each particle. The system corrects the robot arm motion error of the preliminary displacement prediction vector, calculates the centrifugal force offset and Coriolis force interference during the motion process by combining the robot arm kinematic model of the wafer inspection machine, generates a correction coefficient matrix, and weights and fuses the correction coefficient matrix with the preliminary displacement prediction vector to generate the displacement compensation feature quantity of each particle.

7. The method according to claim 6, characterized in that, The process involves using a wafer charge detection unit to acquire charge distribution characteristics on the wafer surface. These characteristics include charge density distribution maps and charge polarity distribution matrices for different regions of the wafer surface. This includes: controlling the probe array of the wafer charge detection unit to perform a gridded scan on the wafer surface, with the scan path moving spirally along the radial direction of the wafer, and the scan step size matching the resolution of the planar coordinate features of the initial particle coordinate set; acquiring charge sensing signals at each scanning sampling point, amplifying the weak sensing signals using a charge amplifier, converting them into digital signals via an analog-to-digital converter, and generating an original charge sensing data set. This original charge sensing data set includes the charge value of each sampling point and the corresponding detection coordinate system coordinates; and processing the original charge sensing data... The dataset undergoes noise suppression by using a wavelet thresholding algorithm to filter out noise from the charge values ​​at each sampling point. Based on the detection coordinates of the original charge sensing data set, spatial interpolation is performed on the denoised charge sensing data to generate a charge density distribution map with continuous spatial distribution. The pixel values ​​in the charge density distribution map represent the charge density feature values ​​at the corresponding locations. A charge polarity distribution matrix is ​​generated based on the sign of the charge density feature values ​​of each pixel in the charge density distribution map. Pixels with positive charge density feature values ​​are marked as first polarity identifiers, pixels with negative charge density feature values ​​are marked as second polarity identifiers, and pixels with zero charge density feature values ​​are marked as third polarity identifiers, thus generating a charge polarity distribution matrix containing charge polarity information.

8. The method according to claim 6, characterized in that, The process of correcting the robot arm motion error of the preliminary displacement prediction vector involves calculating the centrifugal force offset and Coriolis force interference during the motion process using the robot arm kinematic model of the wafer inspection machine, generating a correction coefficient matrix, and then weighted and fused the correction coefficient matrix with the preliminary displacement prediction vector to generate displacement compensation feature quantities for each particle. This includes: acquiring a set of robot arm motion parameters of the wafer inspection machine, which includes the real-time angular velocity values, angular acceleration values ​​of each joint, and the current position coordinates of the robot arm end effector; and, based on the robot arm kinematic model, calculating the centrifugal force acceleration vector and Coriolis force acceleration vector of the robot arm end effector during the motion process according to the set of robot arm motion parameters. The centrifugal force acceleration vector is proportional to the square of the angular velocity of the robot arm joint, and the Coriolis force acceleration vector is proportional to the square of the angular velocity of the robot arm joint. The product of angular velocity and angular acceleration is proportional; the centrifugal force acceleration vector and the Coriolis force acceleration vector are converted into displacement offsets, and the centrifugal force offset and Coriolis force interference of the end effector of the robotic arm are calculated in combination with the robotic arm motion time parameters; an error covariance matrix is ​​constructed based on the centrifugal force offset and Coriolis force interference, and the error covariance matrix is ​​iteratively updated by the Kalman filter algorithm to generate a correction coefficient matrix containing X-axis correction coefficients and Y-axis correction coefficients; the X-axis correction coefficient in the correction coefficient matrix is ​​multiplied by the X-axis displacement component in the initial displacement prediction vector, and the Y-axis correction coefficient is multiplied by the Y-axis displacement component in the initial displacement prediction vector to generate the corrected displacement components, and the corrected X-axis displacement components and Y-axis displacement components are combined to form the displacement compensation feature of each particle.

9. The method according to claim 1, characterized in that, The process of calling the fine-tuning optical system of the wafer inspection machine to perform optical correction response on the displacement compensation feature quantity and generate corrected real-time imaging data includes: analyzing the X-axis and Y-axis displacement components in the displacement compensation feature quantity, generating a correction control command for the fine-tuning optical system, wherein the correction control command includes X-axis and Y-axis fine-tuning parameters, and the units of the fine-tuning parameters are matched with the stepper motor control precision of the fine-tuning optical system; sending the correction control command to the piezoelectric drive unit of the fine-tuning optical system to drive the X-axis and Y-axis translation stages of the fine-tuning optical system to perform micro-displacement adjustments, wherein the displacement adjustment amount of the translation stages is linearly related to the fine-tuning parameters in the correction control command; and calling the laser interferometer of the fine-tuning optical system to perform real-time measurement of the actual displacement of the translation stages. Real-time feedback detection generates a displacement feedback signal, which includes the actual displacement values ​​in the X-axis and Y-axis directions. The displacement feedback signal is compared with the fine-tuning parameters in the correction control command to calculate the displacement adjustment deviation. When the displacement adjustment deviation exceeds a preset deviation threshold, a secondary correction control command is generated to drive the piezoelectric drive unit to perform compensation adjustment until the displacement adjustment deviation is less than the preset deviation threshold. The imaging lens group of the fine-tuning optical system is controlled to adjust the focus. The focus compensation amount is calculated based on the height feature in the initial particle coordinate set and the displacement component in the displacement compensation feature. The focusing motor of the lens group is driven to move axially to generate the imaging optical path after focus adjustment. The corrected imaging sequence after field distortion adjustment and focus drift compensation is collected as real-time imaging data.

10. A computer system, characterized in that, include: A memory storing a computer program; a processor for loading the computer program to implement the vision-based wafer particle position offset detection and compensation method as described in any one of claims 1-9.

Images