Wafer box cleaning state detection method and system based on machine learning

Abstract

The invention relates to the technical field of image recognition, in particular to a wafer box cleaning state detection method and system based on machine learning. The method specifically comprises the following steps: acquiring images of a to-be-detected wafer box and arranging the images into a space angle image set; performing pixel domain deconstruction on each image in the space angle image set to generate a pollutant segmentation image, and identifying and calibrating a pollutant area; according to the pollutant segmentation map, pollution characteristic parameters of each pollutant area are extracted, and cleanliness is evaluated; constructing a three-dimensional geometric model of the wafer box by using the space angle image set; the spatial position of the pollutant area is calculated, the cleanliness and the pollution characteristic parameters are integrated and mapped to the three-dimensional geometric model of the wafer box, a three-dimensional pollution distribution diagram is generated, and a wafer box cleaning state detection report is output. According to the invention, through combination of multi-view three-dimensional reconstruction and deep learning quantitative analysis, quantitative evaluation of spatial distribution of pollutants is realized, and comprehensiveness and accuracy of wafer box cleaning state detection are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, in particular to a wafer box cleaning state detection method and system based on machine learning. BACKGROUND

[0002] After the wafer box is cleaned regularly, the cleaning state of the wafer box needs to be analyzed by image and verified for cleanliness. In the field of wafer box cleaning state image recognition, the existing technology generally uses machine learning-based automatic detection technology to obtain wafer box images, and uses computer vision algorithms (such as threshold segmentation, edge detection, and template matching) to analyze the images, thereby realizing automatic recognition of the cleaning state of the wafer box.

[0003] However, the existing technology has inherent defects in practice. The output result of automatic detection is often a simple "pass / fail" binary judgment, which cannot provide more in-depth analysis information such as specific classification of pollutants, quantitative grading of severity, or pollution distribution rules, thereby failing to adapt to the dynamic changes of the production environment, such as the diversity of pollutant forms, the fluctuation of light conditions, and the aging and wear of the wafer box itself. It is easy to produce blind area, false judgment or omission when detecting the complex three-dimensional images of the inner wall and the card slot of the wafer box.

[0004] Therefore, a wafer box cleaning state detection method and system based on machine learning are proposed. SUMMARY

[0005] The purpose of the present application is to provide a wafer box cleaning state detection method and system based on machine learning to detect the cleaning state of the wafer box.

[0006] To achieve the above purpose, the present application provides the following technical scheme: A wafer box cleaning state detection method based on machine learning, comprising: Collecting images of a wafer box to be tested from multiple spatial perspectives and arranging them into a spatial angle image set; Performing pixel domain deconstruction on each image in the spatial angle image set to generate a pollutant segmentation map; the pollutant segmentation map divides pixels into a background area and a pollutant area, and uses the criterion of eight-neighbor connectivity to mark the boundaries and types of pollutants by breadth-first search; According to the pollutant segmentation map, performing texture pattern extraction on each marked pollutant area, calculating pollutant feature parameters including geometric pattern and image texture, and inputting them into a cleanliness evaluation regression model to evaluate the cleanliness of the corresponding pollutant area according to expert evaluation cleanliness grade; With the spatial angle image set, feature point matching is performed and a polygonal curved surface is fitted to construct a wafer box three-dimensional geometric model; the spatial position of the contaminant region is calculated, the corresponding cleanliness and contaminant characteristic parameters of the contaminant region are integrated, and a three-dimensional contaminant distribution map is generated by mapping the wafer box three-dimensional geometric model to output a wafer box cleaning state detection report.

[0007] Preferably, the specific implementation process of arranging the images of the wafer box to be measured from multiple spatial visual angles into a spatial angle image set includes: adopting a multi-spectral imaging technology, rotating around the wafer box to be measured in multiple axes, and capturing wafer box surface images from different azimuths and pitch angles; all the two-dimensional images collected together constitute the spatial angle image set of the wafer box to be measured.

[0008] Preferably, the specific implementation process of performing pixel domain deconstruction on each image in the spatial angle image set to generate a contaminant segmentation map includes: inputting the image to be processed into a deep learning segmentation model, the deep learning segmentation model being learned based on labeled wafer box image data to identify and distinguish wafer box backgrounds and contaminant types; performing feature extraction on the input image layer by layer through convolution operation, performing semantic classification at the pixel level, and assigning a class label to each pixel point; generating a contaminant segmentation map based on the class labels of all pixel points, and identifying the outline, coverage range and belonging category of different contaminant regions in the form of an image mask through different numerical values and color coding.

[0009] Preferably, the specific implementation process of dividing pixels into background regions and contaminant regions by the contaminant segmentation map, and labeling the boundaries and types of the contaminant through breadth-first search based on the criterion of eight-neighbor connectivity includes: aggregating adjacent pixels with the same contaminant type label in the contaminant segmentation map to form a connected region; for each connected region, generating a closed contour line along the peripheral pixels using an edge detection algorithm to define the geometric boundary of the contaminant region in vector coordinates; assigning the contaminant category label shared by all pixels constituting the connected region to the connected region to label the type of the contaminant region.

[0010] Preferably, the specific implementation process of extracting texture and morphology from each labeled contaminant region according to the contaminant segmentation map to calculate contaminant characteristic parameters including geometric morphology and image texture includes: for each contaminant region, quantifying the geometric morphology of the contaminant region by calculating the area, perimeter, circumscribed rectangle and minimum circumscribed circle of the region; statistically analyzing the pixel gray scale in the region, and extracting image texture reflecting the roughness, granularity and directionality of the contaminant surface by using a gray level co-occurrence matrix algorithm; combining the calculated geometric morphology and image texture to form the contaminant characteristic parameters.

[0011] Preferably, the specific implementation process of inputting the cleanliness evaluation regression model to evaluate the cleanliness of the corresponding pollutant area according to the expert evaluation cleanliness grade includes: inputting the pollution characteristic parameters into the cleanliness evaluation regression model; the cleanliness evaluation regression model is constructed based on known pollutant samples and a training data set of corresponding expert evaluation cleanliness grades, and has a nonlinear mapping relationship of cleanliness quantitative indicators; the pollution characteristic parameters are evaluated and operated to output the cleanliness of the corresponding pollutant area.

[0012] Preferably, the specific implementation process of using the spatial angle image set to perform feature point matching and fit a polygonal surface to construct a wafer box three-dimensional geometric model includes: performing feature point matching between images at different viewing angles through a motion recovery structure technology; based on the matching result, the camera pose of each image and the three-dimensional space coordinates of the feature points are calculated synchronously to generate a three-dimensional point cloud; a continuous surface composed of a polygonal mesh is fitted through a surface reconstruction algorithm to construct a three-dimensional geometric model of the wafer box to be measured.

[0013] Preferably, the specific implementation process of calculating the spatial position of the pollutant area, integrating the corresponding cleanliness and pollution characteristic parameters of the pollutant area, mapping to the wafer box three-dimensional geometric model to generate a three-dimensional pollution distribution map, and outputting a wafer box cleaning state detection report includes: converting the pixel coordinates of the pollutant area in the two-dimensional image to the wafer box three-dimensional geometric model through reverse projection calculation to locate the position and boundary of each pollutant; using a visual coding rule, the three-dimensional model is rendered to obtain a three-dimensional pollution distribution map on the surface of the wafer box; integrating the parameter list of the spatial position, cleanliness and pollution characteristic parameters of each pollutant area in the three-dimensional pollution distribution map, and outputting a wafer box cleaning state detection report.

[0014] A wafer box cleaning state detection system based on machine learning includes: An image acquisition module acquires images of a wafer box to be measured from multiple spatial viewing angles and arranges them into a spatial angle image set; A pixel segmentation module performs pixel domain decomposition on each image in the spatial angle image set to generate a pollutant segmentation map; the pollutant segmentation map divides pixels into background areas and pollutant areas, and uses the criterion of eight-neighborhood connectivity to mark the boundaries and types of pollutants through a breadth-first search method; A cleanliness evaluation module extracts the texture morphology of each marked pollutant area according to the pollutant segmentation map, calculates pollution characteristic parameters including geometric morphology and image texture, and inputs them into a cleanliness evaluation regression model to evaluate the cleanliness of the corresponding pollutant area according to the expert evaluation cleanliness grade; The pollution detection module utilizes the set of spatial angle images, carries out feature point matching and fits a polygonal curved surface, constructs a wafer box three-dimensional geometric model, calculates the spatial position of the pollution area, integrates the cleanliness and pollution characteristic parameters corresponding to the pollution area, maps to the wafer box three-dimensional geometric model to generate a three-dimensional pollution distribution map, and outputs a wafer box cleaning state detection report.

[0015] Compared with the prior art, the present application has the following beneficial effects: 1. The present application can perform omnidirectional wide-area detection on the wafer box without dead angles by collecting images from multiple spatial perspectives and constructing a three-dimensional geometric model, accurately identifies and locates the spatial position, range and severity of all pollutants, and effectively improves the comprehensiveness and accuracy of the detection.

[0016] 2. The present application uses a deep learning model to perform pixel-level segmentation and type identification of the pollutants, and uses a texture morphology extraction and cleanliness evaluation regression model to quantitatively evaluate each pollution area. The final generated three-dimensional pollution distribution map intuitively displays the detailed situation of the pollution, providing deeper data support for the optimization and quality control of the cleaning process.

[0017] 3. The present application uses a machine learning-based method, which enables the system to automatically identify and evaluate pollutants by learning from a large amount of labeled data, reducing manual intervention and avoiding inconsistency caused by subjective judgment. From image acquisition, analysis, three-dimensional modeling to report generation, the entire process is automated, greatly improving the detection efficiency and reliability of the results. BRIEF DESCRIPTION OF DRAWINGS

[0018] Fig. 1 A wafer box cleaning state detection method flow chart based on machine learning is proposed for the present application; Fig. 2 A wafer box cleaning state detection system structure diagram based on machine learning is proposed for the present application; Fig. 3 A wafer box cleaning state detection process schematic diagram is proposed for the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the present application clearer, the following will refer to the drawings and combine specific examples to elaborate the present application in detail. It must be understood that the specific examples described here are only used to explain the present application, and are not intended to constitute any form of limitation on the scope of protection of the present application. Therefore, all equivalent changes or modifications conceived by those skilled in the art based on the content disclosed in the present application without creative labor shall fall within the scope of protection requested by the present application.

[0020] ReferenceFigs. 1 to 3 The application provides a wafer box cleaning state detection method and system based on machine learning. Embodiment one: Referring to Fig. 1 The embodiment provides a wafer box cleaning state detection method based on machine learning, which comprises the following steps. Collect images of the wafer box to be detected from multiple spatial perspectives, and arrange the images into a spatial angle image set. Perform pixel domain decomposition on each image in the spatial angle image set to generate a contaminant segmentation map; the contaminant segmentation map divides pixels into a background region and a contaminant region, and uses the eight-neighborhood connectivity criterion to demarcate the boundary and type of the contaminant through a breadth-first search method. According to the contaminant segmentation map, extract the texture morphology of each demarcated contaminant region, calculate contaminant characteristic parameters including geometric morphology and image texture, and input the parameters into a cleanliness evaluation regression model to evaluate the cleanliness of the corresponding contaminant region according to an expert evaluation cleanliness grade. Use the spatial angle image set to perform feature point matching and fit a polygonal surface to construct a three-dimensional geometric model of the wafer box, calculate the spatial position of the contaminant region, integrate the corresponding cleanliness and contaminant characteristic parameters of the contaminant region, map the parameters to the three-dimensional geometric model of the wafer box to generate a three-dimensional contaminant distribution map, and output a wafer box cleaning state detection report.

[0021] Further, the specific implementation process of collecting images of the wafer box to be detected from multiple spatial perspectives and arranging the images into a spatial angle image set comprises the following steps: using a multi-spectral imaging technology, rotating the wafer box to be detected around multiple axes, and capturing wafer box surface images from different directions and pitch angles; all the two-dimensional images are collectively used to form a spatial angle image set of the wafer box to be detected.

[0022] Specifically, the object to be detected is a 300mm polycarbonate front opening unified pods (FOUP), the core hardware includes an industrial high-resolution CMOS camera with a pixel resolution of 4096x3096 pixels, and a telecentric lens is used to reduce image distortion; a multi-spectral LED light source array controlled by a programmable logic controller, which integrates four groups of light sources with center wavelengths of 450nm (blue light), 530nm (green light), 660nm (red light) and 850nm (near-infrared light); a high-precision industrial robot arm capable of achieving five-axis linkage, used to control the pose of the wafer box to be detected.

[0023] After the acquisition process is started, the end gripper of the industrial robot arm will grab the wafer box to be tested from the conveying track and place it in a closed optical dark box with a diffuse reflection inner wall. The robot arm rotates horizontally around the vertical center line of the wafer box (i.e. azimuth angle) with a step increment of 15 degrees to complete a full 360-degree scan, a total of 24 azimuth angle positions. At each fixed azimuth angle position, the robot arm further controls the wafer box to rotate around the horizontal axis (i.e. pitch angle), for example: from -45 degrees to +45 degrees, with an interval of 15 degrees, to capture 7 different pitch angle poses in turn. Through the combined movement of azimuth angle and pitch angle, the wafer box can be observed from 24 7=168 unique spatial perspectives.

[0024] When the wafer box is stable at each spatial pose, the multi-spectral light source array will be quickly switched and stroboscopically in turn. The CMOS camera is triggered synchronously with the light source, and a two-dimensional gray-scale image is collected under the illumination of each light source (blue, green, red, and near-infrared). Therefore, under each spatial perspective, 4 images of different spectral channels are obtained. After pixel alignment and registration, the 4 images are combined into a four-channel multi-spectral image data unit. After the entire acquisition process is completed, a total of 168 multi-spectral image data units are generated. Through the multi-spectral data dimension, a key physical basis is provided for distinguishing different chemical components of the contaminants. For example: an organic polymer residue that is translucent under visible light shows significant absorption characteristics under 850 nanometer near-infrared spectrum, thereby forming a strong contrast with the polycarbonate substrate, improving the detection sensitivity. The collection of all multi-spectral image data units constitutes a spatial angle image set containing the complete spatial information of the wafer box to be tested.

[0025] This embodiment realizes full coverage of the detection of the complex three-dimensional structure inside the wafer box through multi-axis rotation and multi-view shooting, thereby effectively improving the comprehensiveness and accuracy of the detection. The introduction of multi-spectral imaging technology not only can detect the presence of contaminants, but also can preliminarily classify the materials according to the response differences of the contaminants under different spectra, providing deeper data support for the optimization of cleaning process and the tracing of pollution sources.

[0026] Further, the specific implementation process of pixel domain deconstruction on each image in the spatial angle image set to generate the contaminant segmentation map includes: inputting the image to be processed into a deep learning segmentation model, the deep learning segmentation model is learned based on the labeled wafer box image data, and the wafer box background and the contaminant type are identified and distinguished; the input image is extracted layer by layer through convolution operation, the semantic classification is performed at the pixel level, and the category label is assigned to each pixel point; the contaminant segmentation map is generated based on the category label of all pixel points, and the outline, coverage range and belonging category of different contaminant regions are identified in the form of image mask through different numerical values and color coding.

[0027] Specifically, the deep learning segmentation model used is a U-Net network structure. The training and inference processes of the model are deployed on an industrial computing server equipped with a high-performance graphics processing unit. In the preparation stage of the model, an accurately labeled training data set needs to be constructed. The technical personnel collected more than twenty thousand 300mm wafer box multispectral images of different batches and different contamination levels, and performed fine pixel-level labeling on these images. The labeled categories not only include clean wafer box surfaces (background category), but also subdivide three typical contaminant types commonly seen in semiconductor manufacturing environments, for example: category one is inorganic particulate matter, such as silicon chips and metal particles; category two is organic residues, such as photoresist stains and fingerprint stains; and category three is chemical corrosion spots, such as traces left after acid and alkali reagent splashing. Each original image corresponds to a label image with the same size, and each pixel value in the label image directly corresponds to a category index.

[0028] When a to-be-processed multispectral image selected from the spatial angle image set is input into the trained U-Net model, the image enters the encoder part of the model. The encoder is composed of a series of cascaded convolution layers and pooling layers. The image passes through two consecutive 3 3 convolution layers, and applies a rectified linear unit as an activation function to extract low-order edge, corner and texture features. A 2 2 maximum pooling layer reduces the size of the feature map by half, increases the receptive field and retains the most significant features. This process is repeated four times, and each level of down-sampling enables the model to learn more abstract semantic features. For example: in the deep network, the model can learn that inorganic particulate matter usually presents as small-size blocks with high contrast, while organic residues may present as large-area, edge-blurred low-contrast films.

[0029] The depth feature map output by the encoder is sent to the decoder part of the model, and through a series of up-sampling convolution operations, the size of the feature map is gradually restored to the same size as the original input image. After each up-sampling, it will be spliced with the feature map of the corresponding level of the encoder. The high-resolution, detailed shallow layer features retained in the encoder are fused with the deep layer semantic features recovered by up-sampling in the decoder. This fusion mechanism greatly compensates for the loss of spatial information caused by the pooling operation, enabling the model to accurately understand the category attributes of the pollutants and accurately restore their contour boundaries when performing pixel classification in the end. In the last layer of the decoder, a 1 1 convolutional layer and a Softmax activation function are used to make multi-class predictions for each pixel, outputting probability values for each pixel belonging to each category such as background, inorganic particulate matter, organic residue, or chemical corrosion spot.

[0030] The highest probability value of each pixel is extracted as its final category label, generating a pollutant segmentation map of the same size as the original image, and converting it into a visual image mask. In this mask, different pixel values are assigned different color encodings, with value 0 and black representing the background area, value 1 and red identifying the outline and coverage of all inorganic particulate matter, value 2 and yellow identifying the distribution of organic residues, and value 3 and blue identifying the area of chemical corrosion spots. For example, inputting a wafer box slot image containing tiny silicon chips (diameter about 50 microns) and slight photoresist residues into the model, the model successfully outputs an accurate segmentation map, with the red mask accurately outlining the boundaries of each silicon chip and the yellow mask completely covering the irregularly shaped photoresist residue area.

[0031] This embodiment can automatically and accurately classify multiple pollutants through the powerful feature learning ability of the deep learning model, improving the difficulty of distinguishing different materials and low-contrast pollutants, enhancing the intelligence level and information richness of the detection, and ensuring the high reliability and repeatability of wafer box cleaning quality control.

[0032] Further, the pollutant segmentation map divides pixels into background areas and pollutant areas, and uses the eight-neighborhood connectivity criterion to demarcate the boundaries and types of pollutants through breadth-first search. The specific implementation process includes: aggregating adjacent pixels with the same pollutant type label in the pollutant segmentation map to form a connected region; for each connected region, using an edge detection algorithm to generate a closed contour line along the outer pixels to define the geometric boundary of the pollutant area with vector coordinates; assigning the pollutant category label shared by all pixels constituting the connected region to the connected region to demarcate the type of the pollutant area.

[0033] Specifically, the connected region uses a connected component analysis algorithm to scan each pixel in the pollutant segmentation map line by line. When a non-background category pixel is scanned, this pixel is used as the seed point. Using the eight-neighbor connectivity criterion, a breadth-first search is used to explore and aggregate all pixels that are spatially adjacent to the seed point (i.e., in its eight neighboring positions) and have the same pollutant type label. This aggregation process continues until all pixel members of the connected region are identified and grouped. The originally discrete pixels with the same label are organized into a connected region with inherent integrity.

[0034] After each independent connected region is successfully identified, the Moore's Neighborhood Tracking algorithm is used to explore its outer contour in a clockwise or counterclockwise direction, starting from a boundary pixel of the connected region, and recording the two-dimensional vector coordinates of all pixels that constitute this closed contour line. This ordered set of coordinate points ((x1, y1), (x2, y2), ..., (xn, yn)) digitally defines the geometric boundary of the contaminant region in the image coordinate system. In the detection of the wafer cassette slot region, a contaminant connected region consisting of 128 pixels identified as "metal particles" by the segmentation model is depicted as a closed polygon contour composed of 45 consecutive (x, y) coordinate points through this step. While generating the boundary vector coordinates, the contaminant category label (e.g., label value "1") shared by all pixels constituting the connected region is stored as an attribute of the region.

[0035] This embodiment transforms a disordered set of pixels into bounded, categorized, independent contaminant regions, enabling a refined assessment of contaminant severity. It effectively compresses data volume, facilitating efficient storage and large-scale statistical analysis, and provides a reliable basis for decision-making in tracing contamination sources and evaluating the effectiveness of cleaning processes.

[0036] Furthermore, based on the pollutant segmentation map, the specific implementation process of extracting texture morphology for each labeled pollutant region and calculating pollution feature parameters including geometric morphology and image texture includes: for each pollutant region, quantifying the geometric morphology of the pollutant region by calculating the area, perimeter, bounding rectangle, and minimum bounding circle of the region; performing statistical analysis on the pixel grayscale within the region, and using the gray-level co-occurrence matrix algorithm to extract image texture reflecting the surface roughness, graininess, and directionality of the pollutants; and combining the calculated geometric morphology and image texture to form pollution feature parameters.

[0037] Specifically, the actual coverage area of ​​the contaminant is directly calculated by counting the total number of pixels constituting the contaminant region and combining this with the calibrated ratio of camera pixel size to physical size (e.g., each pixel represents 5 square micrometers). A contaminant region consisting of 314 pixels has a physical area of ​​1.57 square millimeters. Based on the vector coordinates of the closed contour line of this region, its perimeter is calculated by accumulating the Euclidean distance between adjacent coordinate points. A rotating caliper algorithm is used to determine the minimum bounding rectangle and minimum bounding circle that can completely cover the contaminant region. The aspect ratio of the minimum bounding rectangle is used to obtain the shape factor characterizing its extensibility; for example, the shape factor of a long strip of photoresist scratch can reach 8.5, while the shape factor of a near-circular oil droplet is close to 1.1. These values ​​together constitute the set of geometric morphological parameters of the contaminant.

[0038] The pixel grayscale values ​​within a contaminant region (e.g., in the 850 nm near-infrared channel) are treated as independent analysis units. A grayscale co-occurrence matrix (GAM) is constructed to statistically analyze the probability distribution of pixel pairs with different grayscale values ​​co-occurring in different directions (e.g., 0°, 45°, 90°, 135°) and at specific distances (e.g., 1 pixel, 3 pixels). Based on this matrix, texture descriptors are calculated. For example, contaminants with rough surfaces and composed of numerous tiny particles will have significantly higher contrast values ​​(up to 4.8) and lower energy values ​​in their GAM; conversely, smooth and uniform chemically corroded films will have very low contrast values ​​(e.g., 0.5) and homogeneity values ​​very close to 1. These values ​​effectively reflect the microscopic texture of contaminant surfaces, which is difficult for the human eye to consistently discern. All the calculated numerical features are integrated and grouped to generate multi-dimensional contamination feature parameters for each independent contaminant region.

[0039] This embodiment, through numerical description, not only provides an objective and unified standard for pollutant assessment but also lays a solid data foundation for subsequent use of machine learning algorithms to achieve automatic pollutant classification, risk level determination, and source tracing. By performing cluster analysis on a large number of pollution characteristic parameters, precise monitoring and rapid intervention in production processes can be achieved.

[0040] Furthermore, the specific implementation process of inputting into the cleanliness assessment regression model to assess the cleanliness of the corresponding pollutant area according to the cleanliness level rated by experts includes: inputting the pollution characteristic parameters into the cleanliness assessment regression model; the cleanliness assessment regression model is constructed based on the training dataset of known pollutant samples and corresponding expert-rated cleanliness levels, and has a nonlinear mapping relationship with cleanliness quantification indicators; evaluating and calculating the pollution characteristic parameters, and outputting the cleanliness of the corresponding pollutant area.

[0041] Specifically, the cleanliness assessment regression model employs a gradient boosting decision tree architecture. During model construction, over 50,000 representative independent contaminant sample images and their corresponding contamination characteristic parameters are collected from the production line, and the severity of each contaminant sample is independently scored. The expert-rating cleanliness level is a continuous value from 0 to 100, where 100 represents complete cleanliness with no potential risk to production, while 0 represents the most severe contamination requiring immediate rework. Multiple complex factors are considered, including the area and location of the contaminants (e.g., contaminants located in wafer contact areas score significantly lower than similar contaminants in non-critical areas) and their material (e.g., hard metal particles have a much higher risk score than soft organic stains). The cleanliness level of each sample is the average of three expert scores, forming a high-confidence training label set.

[0042] To ensure consistency in scoring, the following criteria table maps the multidimensional characteristics of pollutants to specific deduction items: Baseline score: 100 points.

[0043] Location-based deduction: If the contaminant is located in the slot support area that is in direct contact with the wafer, the base deduction is doubled (e.g., multiplied by a weight of 2.0); if it is located in a non-critical area (outer surface of the wafer box), the deduction is halved (e.g., multiplied by a weight of 0.5).

[0044] Material Deduction: Base deduction values ​​are set based on the type of contaminant (identified by a deep learning segmentation model). For example: Hard metal particles (Category 1) are set as the highest risk due to the potential for wafer scratches, with a base deduction value of -50 points; Chemical corrosion spots (Category 3) are the next highest risk, with a base deduction value of -30 points; Organic residues (Category 2) have the lowest risk, with a base deduction value of -10 points.

[0045] Area / morphology deduction: 5 points will be deducted for every 1 square millimeter increase in the physical coverage area of ​​pollutants; 10 points will be deducted for pollutants whose morphology (defined by the aspect ratio of the smallest bounding rectangle) is elongated (e.g., aspect ratio greater than 5).

[0046] Once the pollution characteristic parameters of a contaminated area to be evaluated are calculated, they are fed into the pre-trained cleanliness assessment regression model. These parameters include values ​​such as an area of ​​1570 square micrometers, a perimeter of 180 micrometers, a shape factor of 1.8, a centroid located in the critical area of ​​the card slot, and a gray-level co-occurrence matrix contrast of 4.8. The cleanliness assessment regression model contains an ensemble structure composed of hundreds of decision trees. This structure represents a nonlinear mapping from the contamination characteristic parameter space to the cleanliness quantification index space learned during training. The input contamination characteristic parameters are sequentially ensembled through these decision trees. In each tree, different feature values ​​are used for a series of logical judgments and path selections, ultimately yielding a basic predicted value at the leaf node. The model then weights and sums the predictions from all trees to obtain the final continuous numerical output. For example, for a contaminant identified as "metal particles," after its contamination characteristic parameters are input into the model, the model performs an evaluation and outputs a cleanliness level of 15.8. A large area of ​​"water stain residue" with extremely low contrast was evaluated by the model, and the cleanliness level was output as 82.3.

[0047] This embodiment extracts complex pollutant characteristic parameters into a single-dimensional cleanliness score, simplifying the decision-making process. It enables the production site to set automated treatment thresholds based on this score, improving production efficiency and automation levels. It also allows companies to quantitatively compare the effects of different cleaning equipment and cleaning solution formulations, thereby continuously improving production processes.

[0048] Furthermore, the specific implementation process of using the spatial angle image set to perform feature point matching and fit polygonal surfaces to construct a three-dimensional geometric model of the wafer box includes: performing feature point matching between images from different perspectives using motion recovery structure technology; synchronously calculating the camera pose and three-dimensional spatial coordinates of feature points for each image based on the matching results to generate a three-dimensional point cloud; and fitting a continuous surface composed of polygonal meshes using a surface reconstruction algorithm to construct a three-dimensional geometric model of the wafer box to be tested.

[0049] Specifically, each image in the spatial angle image set is processed to detect feature points robust to scale, rotation, and illumination changes. In a 4096... In an image with a resolution of 3096 pixels, 3000 to 5000 high-quality feature points can be stably identified. The proposed Structure for Motion technique employs a scale-invariant feature transform algorithm, which establishes an initial correspondence between feature points across all image pairs in the entire image set by performing fast nearest neighbor matching on the feature point descriptors. Based on the two-dimensional feature point matching results, the Structure for Motion technique is initiated to perform bundle adjustment. The system of equations is solved synchronously with the objective function of minimizing the reprojection error of all matched feature points on the image. The solution contains two core parts: first, the accurate camera pose of each image in a unified world coordinate system, including its three-dimensional spatial position (x, y, z) and three-dimensional rotational pose (roll, pitch, yaw); second, the three-dimensional spatial coordinates of all successfully matched feature points. A sparse point cloud consisting of over 150,000 three-dimensional spatial points is generated, which accurately delineates the overall outline and main structural features of the wafer cassette, with its average reprojection error controlled within 0.5 pixels.

[0050] To obtain more refined surface details, the calculated camera pose is further utilized to densify the sparse point cloud using multi-view stereo matching technology, generating a dense point cloud containing millions or even tens of millions of spatial points. This dense point cloud is then used as input, and a Poisson surface reconstruction algorithm is employed to fit an implicit function to the point cloud and extract its isosurfaces, generating a continuous, smooth, and watertight 3D surface. This surface is then constructed as a 3D geometric model of a wafer box composed of over 1.2 million tiny triangular facets. This model reproduces the geometric details of the wafer box under test, including its internal slots, reinforcing ribs, and opening / closing mechanisms.

[0051] This embodiment elevates and unifies local information scattered across numerous independent two-dimensional perspectives into a three-dimensional spatial framework with a true physical scale, providing a coordinate reference for subsequent spatial mapping and analysis of contaminant information. The constructed three-dimensional geometric model of the wafer cassette can analyze the spatial distribution patterns of contaminants, thereby improving the detection accuracy of the wafer cassette cleaning status.

[0052] Furthermore, the specific implementation process of calculating the spatial location of contaminant areas, integrating the cleanliness and contamination characteristic parameters corresponding to the contaminant areas, mapping them to the three-dimensional geometric model of the wafer cassette to generate a three-dimensional contamination distribution map, and outputting a wafer cassette cleaning status detection report includes: calculating the pixel coordinates of contaminant areas in the two-dimensional image through back projection, transforming them into the three-dimensional geometric model of the wafer cassette to locate the position and boundary of each contaminant; rendering the three-dimensional model using visual encoding rules to obtain a three-dimensional contamination distribution map on the surface of the wafer cassette; integrating the parameter list containing the spatial location, cleanliness, and contamination characteristic parameters of each contaminant area in the three-dimensional contamination distribution map, and outputting it as a wafer cassette cleaning status detection report.

[0053] Specifically, for a contaminant region marked in a 2D image (e.g., an image with a 45-degree azimuth and 30-degree pitch angle), the coordinates of all its boundary pixels in the 2D image coordinate system are extracted. Using the camera pose matrix corresponding to the image calculated during the motion reconstruction stage, a 3D spatial ray is constructed for each boundary pixel, originating from the camera's optical center and passing through that pixel, and the intersection of this ray with the surface of the 3D geometric model is calculated. By performing the same back-projection calculation on all boundary pixels of the contaminant region, the 3D outline and coverage of the contaminant region can be delineated on the surface of the 3D model. A region occupying 314 pixels in the 2D image is successfully mapped to a 1.57 square millimeter curved surface on the 3D model surface. To address the issues of repeated detection and inconsistent projection of the same contaminant in multi-view images, a consensus voting mechanism based on a 3D grid is adopted: the surface of the 3D geometric model is divided into a high-density 3D grid (e.g., 0.1 mm). (0.1 mm) Iterates through each image in the spatial angle image set, back-projecting all contaminant regions segmented from it, and marking the 3D raster it covers. For each 3D raster cell, votes are counted to determine how many images from different viewpoints identify (project) the raster cell as a contaminant. A consistency threshold is set (e.g., at least 3 independent viewpoints confirming contamination); only when the cumulative votes for a 3D raster cell exceed this threshold is it finally labeled as a 3D contamination region.

[0054] Once all contaminants identified in the 2D images are consistently mapped to their corresponding spatial locations in the 3D geometric model through this voting mechanism, a visual coding rule is used to render the 3D geometric model itself as a neutral, semi-transparent gray as a background reference. For curved surfaces on the model labeled as contaminants, a color spectrum is applied based on their cleanliness scores. For example, areas with cleanliness scores between 80 and 100 are rendered green, indicating good condition; areas with scores between 50 and 80 are rendered yellow, indicating minor concern; areas with scores between 20 and 50 are rendered orange, indicating moderate contamination; and areas with scores below 20 are rendered in striking red, indicating severe contamination requiring immediate attention. The resulting 3D contamination distribution map allows operators to rotate, zoom, and pan at any angle via an interactive interface, enabling them to know the precise spatial location, impact range, and severity level of each contaminant on the wafer cassette.

[0055] Integrating all information, a detailed wafer cassette cleaning status inspection report is automatically generated. This report is output in a structured data format (e.g., PDF or XML). The first page of the report provides an overall assessment, including the unique identifier of the wafer cassette under test, the inspection time, the total number of contaminants, and the comprehensive cleanliness index calculated as a weighted average of the cleanliness scores of all contaminants. For example, a wafer cassette with a comprehensive cleanliness index of 78.5 is considered "qualified." Subsequent sections of the report list details for each detected contaminant, with each record containing a unique contaminant ID, its centroid coordinates in the 3D model coordinate system, its semantic location (e.g., "inner wall of wafer slot #7"), its physical area, the type of contaminant (e.g., "organic residue"), and its individual cleanliness score.

[0056] This embodiment transforms massive amounts of discrete detection data into a global view through 3D visualization, making it simpler and more efficient to grasp the overall contamination status of the wafer cassette. Precise positioning achieved through back projection provides spatial guidance for subsequent precise cleaning. The generated wafer cassette cleaning status inspection report provides a complete and objective digital archive for wafer cassette quality management, enabling process optimization and yield improvement in semiconductor manufacturing.

[0057] This embodiment improves upon the inherent blind spot problem of fixed single-view detection by acquiring images from multiple spatial perspectives and constructing a spatial angle image set. For the test objects with complex three-dimensional structures of wafer cassettes (e.g., slots, grooves), it can achieve omnidirectional coverage, ensuring the integrity of the original data acquisition, thereby effectively improving the comprehensiveness and accuracy of the detection.

[0058] By utilizing a deep learning segmentation model to deconstruct images at the pixel domain and accurately delineating and classifying contaminants by boundary, detection accuracy is improved. It can depict minute contaminants, ensuring high consistency in detection results.

[0059] By quantifying the geometry, spatial location, and image texture of each contaminated area and inputting this data into a cleanliness assessment regression model to output a precise cleanliness score, the abstract concept of "cleanliness" is transformed into an objective and quantifiable unified indicator. This provides a unified standard for assessing the severity of contaminants, making subsequent work such as tracing contamination sources and evaluating cleaning processes more efficient.

[0060] Example 2: This embodiment provides a machine learning-based wafer cassette cleaning status detection system for wafer cassette cleaning status detection, referring to... Fig. 2 The system includes an image acquisition module, a pixel segmentation module, a cleanliness assessment module, and a contaminant detection module.

[0061] The image acquisition module acquires images of the wafer cell under test from multiple spatial perspectives and organizes them into a spatial angle image set. The pixel segmentation module performs pixel domain deconstruction on each image in the spatial angle image set to generate a pollutant segmentation map. The pollutant segmentation map divides pixels into background regions and pollutant regions. Using the eight-neighbor connectivity criterion and a breadth-first search method, the boundaries and types of pollutants are identified. The cleanliness assessment module extracts the texture and morphology of each marked contaminant area based on the contaminant segmentation map, calculates the contaminant feature parameters including geometric shape and image texture, and inputs them into the cleanliness assessment regression model to assess the cleanliness of the corresponding contaminant area according to the cleanliness level rated by experts. The contaminant detection module uses the spatial angle image set to perform feature point matching and fit polygonal surfaces to construct a three-dimensional geometric model of the wafer box; it calculates the spatial location of the contaminant area, integrates the cleanliness and contamination characteristic parameters corresponding to the contaminant area, maps them to the three-dimensional geometric model of the wafer box to generate a three-dimensional contamination distribution map, and outputs a wafer box cleaning status detection report.

[0062] Furthermore, the image acquisition module employs an integrated system comprising a high-precision industrial robot arm, an industrial-grade high-resolution CMOS camera, and a multispectral LED light source array. The object under test is a 300mm polycarbonate front-opening wafer transfer box. After the acquisition process begins, the end effector of the industrial robot arm grasps the wafer box from the transfer track and places it in a closed optical dark box with diffuse reflection inner walls to eliminate ambient light interference. The robot arm rotates horizontally 360 degrees in 15-degree increments around the wafer box's vertical centerline as its azimuth axis, obtaining 24 different azimuth positions. At each fixed azimuth position, the robot arm further controls the wafer box to rotate around its horizontal axis, i.e., pitch motion, from -45 degrees to +45 degrees, in 15-degree intervals, sequentially capturing seven different pitch angles. Through the combined azimuth and pitch motions, a total of 168 unique spatial perspectives can be observed of the wafer box comprehensively. As the wafer cassette stabilizes at each spatial pose, a multispectral LED light source array rapidly switches and flashes sequentially. This array integrates four light sources with center wavelengths of 450 nm (blue), 530 nm (green), 660 nm (red), and 850 nm (near-infrared). A CMOS camera is triggered synchronously with the light sources, acquiring a two-dimensional grayscale image under each illumination. Therefore, at each spatial viewpoint, the system acquires four images with different spectral channels. After precise pixel alignment and registration, these four images are combined into a four-channel multispectral image data unit. Upon completion of the entire acquisition process, a total of 168 multispectral image data units are generated, collectively forming the spatial angle image set of the wafer cassette under test.

[0063] Furthermore, the pixel segmentation module uses a U-Net deep learning segmentation model, trained to identify and distinguish wafer cassette backgrounds from various contaminant types. During the model preparation phase, technicians collected over 20,000 multispectral images of 300mm wafer cassettes from different batches and with varying levels of contamination, and performed pixel-level annotations on these images. The annotation categories included not only clean wafer cassette surfaces as backgrounds but also three typical contaminant types commonly found in semiconductor manufacturing environments: inorganic particles, organic residues, and chemical corrosion spots. When a multispectral image to be processed is input into the trained U-Net model, the image first enters the encoder section of the model, where it is processed layer by layer through a series of cascaded convolutional and pooling layers, extracting features ranging from low-level features such as edges and corners to more abstract deep semantic features. The depth feature map output by the encoder is fed into the decoder of the model. Through a series of upsampling convolution operations, the size of the feature map is gradually restored to the same level as the original input image. After each upsampling level, it is concatenated and fused with the shallow feature map of the corresponding level of the encoder to compensate for the spatial information loss caused by pooling operations, ensuring that the contour boundaries of the contaminants can be accurately restored. In the last layer of the decoder, a 1 A convolutional layer of size 1 and a softmax activation function perform multi-class prediction for each pixel, outputting the probability value of its belonging to the background or various pollutants. The system extracts the class with the highest probability value for each pixel as its final class label, thereby generating a pollutant segmentation map of the same size as the original image. This segmentation map is further transformed into a visual image mask, where different pixel values ​​are assigned different color codes. For example, values ​​0 and black represent the background area, values ​​1 and red identify the outline and coverage of all inorganic particles, values ​​2 and yellow identify the distribution of organic residues, and values ​​3 and blue identify the area of ​​chemical corrosion spots. The system aggregates adjacent pixels with the same pollutant type label in the pollutant segmentation map to form connected regions. Using a connected component analysis algorithm, each pixel in the segmentation map is scanned row by row. When a pixel that is not in the background category is scanned, this pixel is used as the seed point. Using the eight-neighbor connectivity criterion, a breadth-first search is used to find and aggregate all pixels that are spatially adjacent to the seed point and have the same pollutant type label, until all pixel members of the connected region are identified and grouped. After each independent connected region is successfully identified, the Moore's Neighborhood Tracking algorithm is used to explore its outer contour, starting from a boundary pixel of the connected region, in a clockwise or counterclockwise direction, and record the two-dimensional vector coordinates of all pixels that constitute this closed contour line. This ordered set of coordinate points digitally defines the geometric boundary of the contaminant region in the image coordinate system. Simultaneously, the contaminant category label shared by all pixels constituting the connected region is assigned to that region, completing the type labeling.

[0064] Furthermore, the cleanliness assessment module extracts the texture morphology of each identified contaminated area. For geometric morphology, the actual area covered by the contaminated area is directly calculated by calculating the total number of pixels in the area and combining this with the ratio of the calibrated camera pixel size to the physical size. For example, a contaminated area consisting of 314 pixels can be calculated to have a physical area of ​​1.57 square millimeters. Based on the closed contour vector coordinates of the area, its perimeter is calculated by accumulating the Euclidean distance between adjacent coordinate points, and the minimum bounding rectangle and minimum bounding circle are determined using a rotating caliper algorithm. For image texture, the system treats the pixel grayscale values ​​within the contaminated area as an independent analysis unit. By constructing a grayscale co-occurrence matrix, the probability distribution of pixel pairs with different grayscale values ​​appearing together in different directions and at specific distances within the area is statistically analyzed, and texture descriptors such as contrast, energy, and homogeneity are calculated based on this matrix. These calculated geometric morphology and image texture values ​​together constitute a multi-dimensional set of contamination feature parameters. This parameter set is input into a pre-trained cleanliness assessment regression model, such as a gradient boosting decision tree model. The model is built upon a training dataset of tens of thousands of known contaminant samples and their corresponding expert-rating cleanliness levels (a continuous value from 0 to 100). The model evaluates and calculates contaminant characteristic parameters, ultimately outputting a single-dimensional cleanliness score. For example, a contaminant identified as "metal particles" might have a cleanliness score of 15.8, while a "water stain residue" might have a cleanliness score of 82.3.

[0065] Furthermore, the contaminant detection module utilizes a spatial angle image set and constructs a 3D geometric model of the wafer cassette using motion reconstruction technology. By matching feature points between images from different viewpoints, the camera pose and 3D spatial coordinates of feature points in each image are simultaneously calculated based on the matching results, generating a 3D sparse point cloud. This sparse point cloud is then densified using multi-view stereo matching technology, and a surface reconstruction algorithm is employed to fit a continuous, smooth, and watertight 3D surface composed of millions of tiny triangular facets, thereby constructing the 3D geometric model of the wafer cassette under test. To generate a 3D contamination distribution map, the system back-projects the pixel coordinates of contaminant regions in the 2D image to the 3D geometric model of the wafer cassette using the camera pose matrix calculated in the motion reconstruction stage, thereby locating the position and boundary of each contaminant. For example, a region occupying 314 pixels in a 2D image is successfully mapped to a 1.57 square millimeter surface patch on the 3D model surface. The system utilizes visual coding rules to render and color the corresponding surface patches on the 3D model based on the cleanliness score of each contaminant area. Areas with high scores are rendered in green, and those with low scores are rendered in red. The system integrates a list of parameters such as the spatial location, cleanliness, and contamination characteristics of all contaminants, and outputs a structured wafer cassette cleaning status inspection report.

[0066] This embodiment achieves comprehensive, high-precision, wide-area detection of wafer cassettes by acquiring images from multiple spatial perspectives and constructing a 3D geometric model. This effectively improves the detection blind spot problem caused by a fixed viewpoint, significantly enhancing the comprehensiveness and accuracy of the detection. Deep learning models are used for pixel-level segmentation and type identification of contaminants, and texture morphology extraction and cleanliness assessment regression models are combined to quantitatively evaluate each contaminant region, achieving a leap forward in refined analysis of the specific classification, location, extent, and severity of contaminants. The entire process, from image acquisition, analysis, 3D modeling to report generation, is automated, greatly reducing inconsistencies caused by human intervention and subjective judgment. This not only improves detection efficiency and the reliability of results but also provides deeper data support for the optimization of cleaning processes and quality control.

[0067] Example 3: This embodiment deploys the aforementioned machine learning-based wafer cassette cleaning status detection method and system entirely in the semiconductor workshop of Factory X, referring to... Fig. 3 The cleaning status of the wafer cell is monitored.

[0068] The object under test was a used 300mm polycarbonate front-opening wafer transfer box. This wafer box was gripped on an automated transfer track by a five-axis industrial robotic arm and precisely placed into a closed optical dark box with diffuse-reflective inner walls. The dark box integrated an industrial-grade high-resolution CMOS camera with a pixel resolution of 4096. The camera is equipped with a telecentric lens to eliminate image distortion. A multispectral LED light source array, controlled by a programmable logic controller, is deployed around the camera. This array contains four light sources with center wavelengths of 450 nm (blue light), 530 nm (green light), 660 nm (red light), and 850 nm (near-infrared light). The robotic arm rotates 360 degrees horizontally around the vertical centerline of the wafer cassette as the azimuth axis, in 15-degree increments, to obtain 24 azimuth positions. At each azimuth position, it then rotates around the horizontal axis as the pitch axis, from -45 degrees to +45 degrees, in 15-degree intervals, to capture 7 pitch angles. This combined motion generates a total of 168 unique spatial observation perspectives. At each perspective, the four LED light sources flash sequentially, and the camera synchronously triggers the acquisition of four grayscale images from different spectral channels, ultimately generating a spatial angle image set containing 168 four-channel multispectral image data units.

[0069] For each multispectral image in the spatial angle image set, pixel-domain deconstruction is performed. A deep learning segmentation model based on the U-Net network structure is used to generate a contaminant segmentation map. This model consists of a 5-layer downsampling encoder and a 5-layer upsampling decoder. Each layer of the encoder consists of two consecutive 3 It consists of 3 convolutional layers and a ReLU activation function, followed by 2 A max-pooling layer of size 2 is used for dimensionality reduction. The decoder then uses a 2-level max-pooling layer. The deconvolutional layer 2 is upsampled and concatenated with the feature maps of the corresponding encoder layers to fuse deep semantic information and shallow spatial details. The model is trained on a dataset of 20,000 manually annotated multispectral images. The loss function used during training is a weighted sum of Dice loss and cross-entropy loss. The optimizer is Adam, with an initial learning rate of 0.001 and a learning rate decay strategy. The batch size is set to 4, and the training lasts for 200 epochs. To enhance the model's generalization ability, data augmentation techniques such as random rotation, horizontal flipping, and brightness adjustment are applied during training. For example, when an image containing the 12th slot area of ​​the wafer cassette is input into the model, the model outputs a segmentation mask the same size as the original image, where pixel values ​​0 represent the background, 1 represent inorganic particles (such as silicon chips), 2 represent organic residues (such as photoresist residue), and 3 represent chemical etching spots.

[0070] Based on the contaminant segmentation map, texture morphology extraction and cleanliness assessment are performed on each identified contaminant region. For each independent contaminant region, contaminant characteristic parameters consisting of 12 dimensions are calculated. These parameters specifically include: area measured in physical units (square micrometers), aspect ratio characterizing its extensibility, radius of the smallest circumscribed circle, average pixel grayscale value in the 850 nm near-infrared channel, and texture descriptors such as contrast, energy, and homogeneity calculated through the gray-level co-occurrence matrix (GLCM). For example, a contaminant identified as "metallic particles" has a calculated area of ​​1570 square micrometers, an aspect ratio of 1.2, an average near-infrared channel grayscale of 35 (within the range of 0-255), and a GLCM contrast ratio of 4.8. This 12-dimensional feature vector was input into a cleanliness assessment regression model, which employed a gradient boosting decision tree algorithm. The key hyperparameters were set as follows: 300 decision trees, a maximum depth of 5, and a learning rate of 0.01. The model was constructed based on a training set of tens of thousands of known contaminant samples and their corresponding scores (0-100). After model evaluation, the contamination feature parameters for "metal particles" resulted in a cleanliness score of 15.8, indicating severe contamination and high risk. In contrast, the contamination feature parameters for "water stains" showed a large area but extremely low contrast (0.5), resulting in a cleanliness score of 82.3 after model evaluation, indicating a lower risk.

[0071] A scale-invariant feature transform algorithm was used to detect and describe feature points in all 168 images, and cross-image feature point correspondences were established through fast nearest neighbor matching. Based on these matching results, a clustering adjustment algorithm was used to simultaneously optimize and calculate the 3D pose (position and rotation matrices) of all cameras and a sparse 3D point cloud consisting of over 150,000 points. Multi-view stereo matching technology was used to densify this sparse point cloud, generating a dense point cloud containing millions of points. A Poisson surface reconstruction algorithm (with an octree depth parameter set to 10) was used to fit this dense point cloud, generating a watertight 3D geometric model of a wafer cassette consisting of over 1.2 million triangular facets. An iterative nearest-point algorithm was used to align the model with the standard CAD digital model of the wafer cassette and unify the coordinate system. For each contaminant detected in the 2D image, using the calculated camera pose matrix, back projection was used to accurately map all its boundary pixels in the 2D image onto the surface of the 3D model, thus outlining the location, contour, and coverage of the contaminant in 3D space.

[0072] The system integrates the spatial location, cleanliness level, and contamination characteristic parameters of all contaminants to generate a 3D contamination distribution map and output a wafer cassette cleaning status inspection report. In this 3D distribution map, the wafer cassette model itself is rendered in semi-transparent gray, while the contaminants mapped onto its surface are color-coded according to their cleanliness scores: areas with scores below 20 are rendered in red (severely contaminated), 20 to 50 in orange, 50 to 80 in yellow, and above 80 in green (good condition). Users can freely rotate and zoom the 3D model through an interactive interface to intuitively view the details of each contaminant. The final generated PDF inspection report shows the wafer cassette ID, inspection time, and a comprehensive cleanliness index (e.g., 78.5, judged as "qualified") on the first page. Subsequent pages record detailed information for each detected contaminant in a list format, including its unique contaminant ID, its spatial location (e.g., "inner wall of wafer slot #7, 12 mm from the bottom support point"), physical area, contaminant type, and cleanliness level, enabling the detection of the wafer cassette cleaning status within the X factory semiconductor workshop.

[0073] It should be clarified that the embodiments described above are merely exemplary and are intended to aid in understanding the present invention, not to limit it. Those skilled in the art can make various changes and modifications after grasping the core ideas of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A wafer cassette cleaning status detection method based on machine learning, characterized in that, include: Images of the wafer cell under test were acquired from multiple spatial perspectives and compiled into a spatial angle image set; Each image in the spatial angle image set is deconstructed in the pixel domain to generate a pollutant segmentation map; the pollutant segmentation map divides the pixels into background regions and pollutant regions, and uses the eight-neighbor connectivity criterion and breadth-first search to mark the boundaries and types of pollutants. Based on the pollutant segmentation map, texture morphology is extracted for each marked pollutant region, and pollutant feature parameters including geometric morphology and image texture are calculated and input into the cleanliness assessment regression model. The cleanliness of the corresponding pollutant region is assessed according to the cleanliness level rated by experts. Using the spatial angle image set, feature points are matched and polygonal surfaces are fitted to construct a three-dimensional geometric model of the wafer box; the spatial location of the contaminant area is calculated, and the cleanliness and contamination characteristic parameters corresponding to the contaminant area are integrated and mapped to the three-dimensional geometric model of the wafer box to generate a three-dimensional contamination distribution map, and a wafer box cleaning status detection report is output.

2. The wafer cassette cleaning status detection method based on machine learning according to claim 1, characterized in that, The specific process of acquiring images of the wafer cassette under test from multiple spatial perspectives and organizing them into a spatial angle image set includes: using multispectral imaging technology, rotating around the wafer cassette under test along multiple axes to capture images of the wafer cassette surface from different orientations and pitch angles; all the acquired two-dimensional images together constitute the spatial angle image set of the wafer cassette under test.

3. The wafer cassette cleaning status detection method based on machine learning according to claim 1, characterized in that, The specific implementation process of generating a contaminant segmentation map by deconstructing each image in the spatial angle image set into the pixel domain includes: inputting the image to be processed into a deep learning segmentation model, which learns based on labeled wafer cassette image data to identify and distinguish the wafer cassette background from the contaminant type; extracting features from the input image layer by layer through convolution operations, performing semantic classification at the pixel level, and assigning a category label to each pixel; generating a contaminant segmentation map based on the category labels of all pixels, and identifying the contour, coverage area, and category of different contaminant regions in the form of an image mask through different numerical values ​​and color codes.

4. The wafer cassette cleaning status detection method based on machine learning according to claim 1, characterized in that, The pollutant segmentation map divides pixels into background and pollutant regions. Using an eight-neighbor connectivity criterion and a breadth-first search approach, the specific implementation process for defining the boundaries and types of pollutants includes: aggregating adjacent pixels with the same pollutant type label in the pollutant segmentation map to form connected regions; for each connected region, using an edge detection algorithm to generate closed contour lines along the outer pixels, defining the geometric boundaries of the pollutant region using vector coordinates; and assigning the pollutant category label shared by all pixels constituting the connected region to the connected region, thus defining the type of the pollutant region.

5. The wafer cassette cleaning status detection method based on machine learning according to claim 1, characterized in that, Based on the pollutant segmentation map, the specific implementation process of extracting texture morphology for each labeled pollutant region and calculating pollution feature parameters including geometric morphology and image texture includes: for each pollutant region, the area, perimeter, bounding rectangle, and minimum bounding circle of the region are calculated to quantify the geometric morphology of the pollutant region; statistical analysis of the pixel grayscale within the region is performed, and the gray-level co-occurrence matrix algorithm is used to extract image texture reflecting the surface roughness, graininess, and directionality of the pollutants; the calculated geometric morphology and image texture are combined to form pollution feature parameters.

6. The wafer cassette cleaning status detection method based on machine learning according to claim 1, characterized in that, The specific implementation process of inputting the cleanliness assessment regression model to evaluate the cleanliness of the corresponding pollutant area according to the cleanliness level rated by experts includes: inputting the pollution characteristic parameters into the cleanliness assessment regression model; the cleanliness assessment regression model is constructed based on the training dataset of known pollutant samples and corresponding expert-rated cleanliness levels, and has a nonlinear mapping relationship with cleanliness quantification indicators; evaluating and calculating the pollution characteristic parameters, and outputting the cleanliness of the corresponding pollutant area.

7. The wafer cassette cleaning status detection method based on machine learning according to claim 1, characterized in that, The specific implementation process of constructing a three-dimensional geometric model of a wafer box by matching feature points and fitting polygonal surfaces using the spatial angle image set includes: matching feature points between images from different perspectives using motion recovery structure technology; synchronously calculating the camera pose and three-dimensional spatial coordinates of feature points for each image based on the matching results to generate a three-dimensional point cloud; and fitting a continuous surface composed of polygonal meshes using a surface reconstruction algorithm to construct a three-dimensional geometric model of the wafer box to be tested.

8. The wafer cassette cleaning status detection method based on machine learning according to claim 1, characterized in that, The specific implementation process of calculating the spatial location of contaminant areas, integrating the corresponding cleanliness and contamination characteristic parameters of the contaminant areas, mapping them to the three-dimensional geometric model of the wafer cassette to generate a three-dimensional contamination distribution map, and outputting a wafer cassette cleaning status detection report includes: calculating the pixel coordinates of contaminant areas in the two-dimensional image through back projection, transforming them into the three-dimensional geometric model of the wafer cassette to locate the position and boundary of each contaminant; rendering the three-dimensional model using visual encoding rules to obtain a three-dimensional contamination distribution map on the surface of the wafer cassette; integrating the parameter list of each contaminant area in the three-dimensional contamination distribution map, including spatial location, cleanliness, and contamination characteristic parameters, and outputting it as a wafer cassette cleaning status detection report.

9. A wafer cassette cleaning status detection system based on machine learning, characterized in that, include: The image acquisition module acquires images of the wafer cell under test from multiple spatial perspectives and organizes them into a spatial angle image set. The pixel segmentation module performs pixel domain deconstruction on each image in the spatial angle image set to generate a pollutant segmentation map. The pollutant segmentation map divides pixels into background regions and pollutant regions. Using the eight-neighbor connectivity criterion and a breadth-first search method, the boundaries and types of pollutants are identified. The cleanliness assessment module extracts the texture and morphology of each marked contaminant area based on the contaminant segmentation map, calculates the contaminant feature parameters including geometric shape and image texture, and inputs them into the cleanliness assessment regression model to assess the cleanliness of the corresponding contaminant area according to the cleanliness level rated by experts. The contaminant detection module uses the spatial angle image set to perform feature point matching and fit polygonal surfaces to construct a three-dimensional geometric model of the wafer box; it calculates the spatial location of the contaminant area, integrates the cleanliness and contamination characteristic parameters corresponding to the contaminant area, maps them to the three-dimensional geometric model of the wafer box to generate a three-dimensional contamination distribution map, and outputs a wafer box cleaning status detection report.

Images