Intelligent detection method and system for detection device

Abstract

The application discloses an intelligent detection method and system of a detection device, and relates to the technical field of intelligent detection. Different dimensional data combinations are determined according to analysis of multi-modal data, corresponding first defect areas are determined according to identification of each data combination, the first defect areas are dynamically coupled, and primary defect information of a silicon wafer is further determined in combination with a constraint relationship of the silicon wafer, thereby improving the accuracy of the primary defect information of the silicon wafer. A second defect area to be reviewed is determined according to negative feedback of defect evolution information, corresponding three-dimensional defect features are further determined in combination with the first defect areas, a plurality of production processes of the silicon wafer are marked, corresponding abnormal processing nodes are determined in combination with the three-dimensional defect features of the silicon wafer, a plurality of abnormal processing information corresponding to the silicon wafer is determined according to tracing of the abnormal processing nodes, and supplementary detection of the detection device is further triggered, thereby improving the accuracy of the intelligent detection sequence.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, and more particularly to an intelligent detection method and system for a detection device. Background Technology

[0002] With the continuous evolution of semiconductor manufacturing processes, ultra-thin silicon wafers have become the mainstream in the industry. Ultra-thin silicon wafers, such as 12-inch single-crystal silicon wafers with a thickness of less than 100 micrometers, have a very large specific surface area and extremely low stiffness. When macroscopic defects are detected on a vacuum adsorption stage, uncontrollable local warping and high-frequency micro-vibrations will inevitably occur.

[0003] Existing inspection devices typically use fixed light sources and shooting parameters, which cannot be dynamically adjusted according to the real-time changes in the posture data of the silicon wafer. When the silicon wafer warps or vibrates, the camera often captures images that are locally overexposed, shadowed, or blurred due to defocus. Defect judgment is based solely on a single two-dimensional image. This single-dimensional inspection mode cannot decouple the real abnormal signals from complex background noise. The initial defect area is fragmented and contains a large number of false defects, which affects the accuracy of primary defect information and leads to low accuracy of intelligent inspection sequences. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides an intelligent detection method and system for a detection device.

[0005] This invention provides an intelligent detection method for a detection device, comprising: During the inspection of silicon wafers by the inspection device, the composite imaging module of the inspection device is adjusted according to the orientation data of the silicon wafer to synchronously acquire real-time images of the silicon wafer, and multimodal data is determined by combining the production process of the silicon wafer and the corresponding surface texture. Based on the analysis of the multimodal data, multiple data combinations of different dimensions are determined. Based on the identification of each data combination, the corresponding first defect region is determined. Multiple first defect regions are dynamically coupled, and the primary defect information of the silicon wafer is further determined by combining the constraint relationship of the silicon wafer. The primary defect information is input into the digital twin space and dynamically extrapolated using a time-series extrapolation mechanism, thereby outputting the defect evolution information of the silicon wafer. Based on the negative feedback of this defect evolution information, the second defect area to be re-examined is determined, and the corresponding three-dimensional defect features are further determined by combining multiple first defect areas. Multiple production processes of silicon wafers are marked, and the corresponding abnormal processing nodes are determined by combining the three-dimensional defect characteristics of the silicon wafers. Based on the tracing of the abnormal processing nodes, multiple abnormal processing information corresponding to the silicon wafers are determined, which further triggers supplementary detection by the detection device. A closed-loop intelligent detection sequence is constructed by combining the current working status of the detection device.

[0006] This invention provides an intelligent detection system for a detection device, which is applied to the intelligent detection method described above; the intelligent detection system includes: The multimodal data module is used to adjust the composite imaging module of the inspection device according to the posture data of the silicon wafer during the inspection process of the inspection device, synchronously acquire real-time images of the silicon wafer, and determine multimodal data by combining the production process of the silicon wafer and the corresponding surface texture. The primary defect information module is used to determine multiple data combinations of different dimensions based on the parsing of the multimodal data, determine the corresponding first defect region based on the identification of each data combination, dynamically couple multiple first defect regions, and further combine the constraint relationship of the silicon wafer to determine the primary defect information of the silicon wafer. The defect region module is used to input primary defect information into the digital twin space and perform dynamic simulation in combination with the time-series simulation mechanism, thereby outputting the defect evolution information of the silicon wafer. Based on the negative feedback of the defect evolution information, the second defect region to be re-examined is determined, and the corresponding three-dimensional defect features are further determined by combining multiple first defect regions. The intelligent detection module is used to mark multiple production processes of silicon wafers, determine the corresponding abnormal processing nodes by combining the three-dimensional defect characteristics of silicon wafers, determine multiple abnormal processing information corresponding to silicon wafers by tracing the abnormal processing nodes, further trigger supplementary detection by the detection device, and construct a closed-loop intelligent detection sequence by combining the current working status of the detection device.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) During the inspection of silicon wafers by the inspection device, the composite imaging module of the inspection device is adjusted according to the posture data of the silicon wafer to synchronously acquire real-time images of the silicon wafer, and multimodal data is determined by combining the production process of the silicon wafer and the corresponding surface texture; multiple data combinations of different dimensions are determined by analyzing the multimodal data, and the corresponding first defect area is determined by identifying each data combination. Multiple first defect areas are dynamically coupled, and the primary defect information of the silicon wafer is determined by combining the constraint relationship of the silicon wafer. Multimodal data is introduced to control the first defect area and improve the accuracy of the primary defect information of the silicon wafer.

[0008] (2) Input the primary defect information into the digital twin space and perform dynamic simulation in combination with the time-series simulation mechanism to output the defect evolution information of the silicon wafer. Based on the negative feedback of the defect evolution information, determine the second defect area to be re-inspected. Further combine multiple first defect areas to determine the corresponding three-dimensional defect features. Mark multiple production processes of the silicon wafer and determine the corresponding abnormal processing nodes in combination with the three-dimensional defect features of the silicon wafer. Based on the tracing of the abnormal processing nodes, determine multiple abnormal processing information corresponding to the silicon wafer. Further trigger the supplementary detection of the detection device and construct a closed-loop intelligent detection sequence in combination with the current working state of the detection device to further control the second defect area. Fully consider multiple abnormal processing information and the current working state of the detection device to improve the accuracy of the intelligent detection sequence. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the intelligent detection method of the detection device in this embodiment of the invention; Figure 2 This is a schematic diagram of the structural composition of the intelligent detection system of the detection device in this embodiment of the invention. Detailed Implementation

[0010] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0011] Please see Figures 1 to 2 An intelligent detection method for a detection device, applied in intelligent detection scenarios; the intelligent detection method for the detection device includes: Step S11: During the inspection of the silicon wafer by the inspection device, the composite imaging module of the inspection device is adjusted according to the orientation data of the silicon wafer to synchronously acquire real-time images of the silicon wafer, and multimodal data is determined by combining the production process of the silicon wafer and the corresponding surface texture. Step S12: Based on the analysis of the multimodal data, determine multiple data combinations of different dimensions, determine the corresponding first defect region based on the identification of each data combination, dynamically couple multiple first defect regions, and further combine the constraint relationship of the silicon wafer to determine the primary defect information of the silicon wafer. Step S13: Input the primary defect information into the digital twin space and perform dynamic simulation in combination with the time-series simulation mechanism to output the defect evolution information of the silicon wafer. Based on the negative feedback of the defect evolution information, determine the second defect area to be re-examined, and further combine multiple first defect areas to determine the corresponding three-dimensional defect features. Step S14: Mark multiple production processes of the silicon wafer, determine the corresponding abnormal processing nodes by combining the three-dimensional defect characteristics of the silicon wafer, determine multiple abnormal processing information corresponding to the silicon wafer by tracing the abnormal processing nodes, further trigger the supplementary detection of the detection device, and construct a closed-loop intelligent detection sequence by combining the current working status of the detection device.

[0012] In step S11, the specific steps are as follows: S111: When the inspection device performs silicon wafer inspection tasks, it acquires the three-dimensional posture data of the silicon wafer at the inspection station in real time based on the posture sensor. It determines multiple inspection directions based on the recognition of the three-dimensional posture data of the silicon wafer, and uses the three-dimensional posture data of the silicon wafer as a feedforward signal to trigger the control of the composite imaging module of the inspection device, so that the composite imaging module of the inspection device can adaptively inspect the silicon wafer along different angles to acquire real-time images of the silicon wafer. S112: Obtain the production process and surface texture of the silicon wafer, align the surface texture and real-time image of the silicon wafer in time and space, dynamically construct the corresponding data framework, and further combine the production process of the silicon wafer to perform multi-factor fusion, thereby constructing multimodal data with high-order correlation.

[0013] In the embodiments of this application, when the detection device performs the silicon wafer detection task, it acquires the three-dimensional posture data of the silicon wafer at the detection station in real time based on the pose sensor. Multiple detection directions are determined based on the identification of the three-dimensional posture data of the silicon wafer. The three-dimensional posture data of the silicon wafer is used as a feedforward signal to trigger the control of the composite imaging module of the detection device, so that the composite imaging module of the detection device adaptively detects the silicon wafer along different angles to acquire real-time images of the silicon wafer. This takes into account the overall consideration of the three-dimensional posture data of the silicon wafer and ensures the accuracy of multiple detection directions.

[0014] At this time, when the inspection device performs the silicon wafer inspection task, the system uses an array of pose sensors deployed around and on top of the inspection station to collect the spatial coordinate sequence of the silicon wafer in real time with a high sampling frequency in a non-contact manner, thereby determining the six-degree-of-freedom three-dimensional attitude data including translational offset and three-dimensional rotation angle. During this process, the pose sensors capture the relative displacement of the silicon wafer under vacuum adsorption or mechanical transmission, and convert the original spatial point cloud signal into an attitude matrix with a physical topology in real time, providing a reference spatial coordinate system for subsequent optical tracking.

[0015] The system inputs the received 3D pose data into a pre-built orientation decision model for feature recognition. Based on the current tilt vector, curvature distribution gradient, and surface normal deflection state of the silicon wafer, the model determines multiple spatial incident vectors that can eliminate geometric occlusion and optical distortion, and establishes multiple directions to be detected. This recognition process is not an exhaustive search of fixed angles, but a dynamic calculation based on the actual pose of the current silicon wafer, ensuring that the generated directions to be detected can cover the potential feature exposure angles of the silicon wafer surface to the greatest extent in 3D space, thereby breaking through the physical blind spot of a single vertical detection perspective.

[0016] The system directly injects the calculated 3D attitude data as a feedforward signal into the underlying motion controller of the detection device. Before the silicon wafer fully enters a static stable state, it triggers the pose compensation and multi-channel optical control of the composite imaging module in advance. After receiving the feedforward command, the multi-axis servo mechanism inside the composite imaging module works in coordination with the variable focal length optical lens group to perform trajectory interpolation and angle deflection in 3D space along the multiple directions to be detected as determined above. This achieves 3D adaptive decoupling control of the light source incident angle, camera line of sight pitch angle, and focal plane compensation, so that the optical observation link always maintains the best matching relative geometric relationship with the local morphology of the silicon wafer.

[0017] During the dynamic tracking process of adaptive angle deflection completed by the composite imaging module, the multispectral light source and image sensor inside the composite imaging module perform high-frequency synchronous exposure according to the preset timing logic, and continuously scan the silicon wafer in motion or micro-floating state along different angles, thereby acquiring real-time images of the silicon wafer that eliminate defocus blur and shadow occlusion caused by pose tilt. These real-time images not only contain surface reflectivity distribution information from multiple perspectives, but also implicitly contain three-dimensional morphology based on the current dynamic perspective, directly providing a high signal-to-noise ratio and multi-dimensional complementary original visual data source for subsequent high-precision defect analysis.

[0018] Specifically, the detection device is a defect detection device equipped with a high-speed camera, which inspects the silicon wafer. When the detection device starts to perform the inspection task of the silicon wafer, due to the extreme ultra-thin characteristics of the silicon wafer, it undergoes a warping deformation with the edges bending downward by hundreds of micrometers after being adsorbed on the stage. At this time, the laser displacement and pose sensor built into the detection device scans the silicon wafer in real time and obtains three-dimensional pose data of its edge region normal vector being seriously deviated from the vertical direction.

[0019] Based on the recognition of this attitude data, the control system of the detection device determines that the traditional vertical top-down detection direction will be completely interfered with by the reflection of the sloping surface of the warped area. Therefore, it generates several targeted detection directions, such as "25-degree left tilt dark field incidence" and "25-degree right tilt bright field incidence". Without waiting for the stage to come to a complete stop, the system directly sends the three-dimensional attitude data containing the warping slope and displacement vector as a feedforward signal to the composite imaging module. After receiving the feedforward signal, the tilt servo mechanism of the composite imaging module performs adaptive deflection and tracking along the determined 25-degree left and 25-degree right tilt directions while the silicon wafer rotates at high speed with the stage into the field of view. At the same time, the Z-axis fast response mechanism performs focal plane compensation synchronously according to the height difference of the warped surface. During the process of the composite imaging module adaptively tracking the warped edge of the silicon wafer at different angles, the image sensor simultaneously completes multi-angle exposure and successfully acquires clear and shadow-free real-time images of the warped sloping surface of the silicon wafer.

[0020] Furthermore, the production process and surface texture of the silicon wafer are obtained, and the surface texture and real-time image of the silicon wafer are spatiotemporally aligned to dynamically construct the corresponding data framework. In addition, the production process of the silicon wafer is further combined with multi-factor fusion to construct multimodal data with high-order correlation, which is compatible with the overall consideration of the silicon wafer production process and ensures the accuracy of multimodal data with high-order correlation.

[0021] At this point, the system retrieves the production history of the silicon wafer throughout its entire lifecycle through the data interface of the underlying manufacturing execution system. This production history is structured and parsed into a time-series feature vector containing multiple machining parameters, consumable batch attributes, and environmental stress. Meanwhile, the system extracts the inherent surface texture of the silicon wafer through pre-calibration or pre-scanning. This texture represents the microscopic physical morphology of the silicon wafer with specific spatial frequencies and geometric topological laws left over from the crystal cutting or grinding stage.

[0022] The system maps the extracted surface textures to the visual inspection coordinate system and performs strict spatiotemporal alignment with the multi-angle real-time images acquired in the previous stage. In the spatial dimension, based on the three-dimensional pose data of the silicon wafer, the physical coordinates of the surface textures are registered with the image pixel coordinates through an affine transformation to eliminate projection distortion caused by viewpoint deflection. In the temporal dimension, using the exposure timestamp of the image sensor as a reference, the dynamically moving surface textures of the silicon wafer are cut into local texture slices that correspond one-to-one with a single frame of real-time image. The system uses this spatiotemporal alignment result as a skeleton to dynamically construct a multi-dimensional data framework that nests a pixel intensity matrix, a texture spatial distribution matrix, and relative time labels.

[0023] The system injects the previously parsed production process time-series feature vector into the dynamic data framework and performs a deep multi-factor fusion operation. In this process, specific process parameters in the production process are used as prior conditions to perform cross-correlation analysis on the image pixel features and texture features within the data framework. This process separates the regular texture response caused by normal processing and the random mutation response caused by abnormal defects. By combining low-dimensional visual data with high-dimensional process causal data in this way, the system finally outputs a multimodal data with high-order correlation that breaks down the barriers between heterogeneous data. This allows the final data representation to directly reflect the physical root cause of the defect rather than simply the apparent optical differences.

[0024] Optionally, for cross-correlation analysis, the system first extracts two types of core features from the constructed dynamic data framework: one type is pixel-level features derived from real-time images, including the grayscale values ​​of each pixel and their gradient changes under multi-angle light source illumination; the other type is texture features derived from surface texture analysis, including geometric parameters such as the local orientation angle, spatial frequency, and groove spacing of the texture. These two types of features have already established a correspondence on pixel coordinates based on spatiotemporal alignment.

[0025] The system incorporates specific process parameters from the production process as prior conditions. These parameters represent a quantitative characterization of the mechanical or thermal effects experienced by the silicon wafer during a particular processing step, such as the grit size, feed rate, and local pressure distribution in a grinding process. Based on these process parameters and a known process physics model, the system pre-calculates the optical response range or trend that a normal texture structure should produce under the current process conditions. This pre-calculation does not rely on fixed empirical formulas but is based on an understanding of the causal relationship between processing technology and surface morphology; that is, specific processing methods leave characteristic textures with definite spatial regularities on the silicon wafer surface, and these textures produce predictable optical response patterns under specific lighting conditions.

[0026] After completing the aforementioned prior modeling, the system performs a pixel-by-pixel correlation and comparison between the pixel features extracted from the real-time image and the theoretical optical response corresponding to the texture features at that pixel location. Specifically, the system calculates the deviation between the measured pixel features and the theoretical prediction values. If the deviation is within the allowable range, it indicates that the optical signal reflected by the pixel conforms to the regular texture response caused by normal processing technology, and the system classifies it as a background signal and retains it without marking it as an anomaly. If the deviation significantly exceeds the allowable range, it indicates that the pixel has a sudden change that cannot be explained by normal process rules, and the system classifies it as an abnormal sudden change response caused by defects, extracts it, and marks it. Through the aforementioned pixel-by-pixel cross-correlation residual analysis, the system achieves the function of accurately separating abnormal signals from complex optical backgrounds. The core of this process is that the system does not rely on fixed grayscale or contrast thresholds to determine defects, but uses process rules as a dynamic reference benchmark, thereby enabling the judgment results to adapt to batch-to-batch process fluctuations.

[0027] Specifically, after the detection device acquires a real-time image of the warped edge of the silicon wafer through the composite imaging module, the system obtains the production history of the silicon wafer from the upstream MES system and extracts the grinding wheel grit, feed speed and local pressure distribution curve of its last "back rough grinding" process; at the same time, the system obtains the surface texture of the silicon wafer, which is inevitably left by the #800 grinding wheel and is presented as concentric circular grooves with a specific spacing.

[0028] Because silicon wafers exhibit high-frequency micro-vibrations during inspection, there is a dynamic misalignment between the real-time image and the surface texture at the pixel level. Therefore, the system utilizes the micro-vibration frequency and three-dimensional attitude offset calculated in stage S111 to perform millisecond-level spatiotemporal alignment between the projected coordinates of the concentric circular surface texture and the pixel coordinates of the real-time image captured by the high-speed camera of the inspection device. This dynamically constructs a dynamic data framework that includes "pixel grayscale values ​​under tilted viewpoints - depth distribution of the registered concentric circular grooves - micro-vibration phase labels".

[0029] Within this framework, the system begins to perform multi-factor fusion: due to the extreme thinness of the silicon wafer, its warped edges, under the illumination of the multi-angle annular bright and dark field light source of the detection device, will produce strong, regular scattered light spots from the normal concentric circle texture; the system incorporates the production process factor of "grinding wheel feed speed" into the fusion method, thereby learning "what specific gradient distribution should the scattered light spots produced by the concentric circle texture exhibit at the current feed speed". Through this high-order correlation calculation, light spots in the real-time image that conform to the regular gradient distribution are judged as normal texture optical responses and filtered out, while abnormal reflective points that deviate from the regular gradient are accurately identified; the system deeply binds optical image features, physical texture features and rough grinding process parameters to construct multimodal data with high-order correlation that can penetrate the complex optical interference on the silicon wafer surface; optional, abnormal reflective points include, for example, microcracks caused by local stress release.

[0030] Optionally, the system first establishes a process-optical response benchmark model based on the parameters of the rough grinding process. The core function of this benchmark model is to calculate the light spot distribution pattern that a normal concentric circle pattern on the silicon wafer surface should exhibit under multi-angle annular bright and dark field light source illumination, based on given process conditions. This calculation is not a simple empirical fit, but is based on the mechanism of the effect of process parameters on surface morphology, as well as the influence of surface morphology on optical reflection and scattering, thereby obtaining theoretical light spot gradient distribution characteristics with physical interpretability.

[0031] Based on this, the system extracts gradient features from each light spot detected in the real-time image, including the period of grayscale variation along the concentric circle tangent, the attenuation slope along the radial direction, and the relative ratio of scattering intensity under different light source angles, among other multi-dimensional gradient information. Subsequently, the system performs similarity matching between the measured gradient feature vector and the theoretical gradient feature vector calculated by the benchmark model. This similarity matching is a comprehensive comparison across multiple dimensions, rather than a simple comparison of a single indicator.

[0032] When the degree of matching between the measured gradient features and the theoretical gradient features reaches the preset similarity requirement, the system determines that the generation of the light spot conforms to normal process rules and belongs to the normal scattering response of concentric circle textures to the light source. Therefore, it is filtered out as background noise and not included in the defect candidate set. Conversely, when there is a significant deviation between the measured gradient features and the theoretical gradient features, the system determines that the generation of the light spot cannot be explained by normal texture scattering rules and belongs to an abnormal reflective point. It is then precisely stripped and marked as a suspected defect area that requires further analysis.

[0033] The key to the aforementioned high-order correlation calculation lies in the system's ability to quantitatively predict normal optical responses using process parameters as priors. This enables a physical-based distinction between "normal" and "abnormal" responses, rather than relying on subjectively set fixed thresholds for coarse classification. This allows the system to maintain a stable ability to extract abnormal signals when faced with silicon wafers from different batches, under different process conditions, or with varying degrees of warpage.

[0034] In step S12, the specific steps are as follows: S121: A manifold learning mechanism is introduced for multimodal data. The multimodal data is dimensionality-reduced and mapped, and multiple data combinations of different dimensions are parsed simultaneously. These multiple data combinations cover the dimensions of morphological features, texture anomalies, and stress distribution. Each data combination is dynamically identified, and multi-scale detection is performed during the identification process to determine multiple first defect regions of the silicon wafer. S122: Load the corresponding graph neural network for multiple first defect regions and trigger the dynamic coupling of multiple first defect regions, thereby transforming multiple first defect regions into nodes in the topology graph. Simultaneously perform feature aggregation on multiple first defect regions, and further combine the constraint relationship of the silicon wafer to perform pseudo-defect elimination and boundary fitting, thereby outputting the primary defect information of the silicon wafer. The primary defect information of the silicon wafer includes the exact location, shape and physical properties of the defects.

[0035] In the embodiments of this application, a manifold learning mechanism is introduced for the multimodal data. The multimodal data is dimensionality-reduced and mapped, and multiple data combinations of different dimensions are parsed simultaneously. These multiple data combinations cover the dimensions of morphological features, texture anomalies, and stress distribution. Each data combination is dynamically identified, and multi-scale detection is performed during the identification process to determine multiple first defect regions of the silicon wafer. This approach is compatible with the overall consideration of each data combination and ensures the accuracy of the multiple first defect regions of the silicon wafer.

[0036] At this point, after acquiring the high-order correlated multimodal data, the system introduces a manifold learning mechanism to perform topology-preserving dimensionality reduction mapping to address the severe nonlinear coupling and high-dimensional redundancy issues within the data. This mechanism projects the multimodal data containing heterogeneous noise into a low-dimensional feature space by finding the intrinsic low-dimensional manifold surface of the data point distribution in the high-dimensional data space. Within this low-dimensional space, based on the clustering boundaries and gradient variance of the feature vectors, it simultaneously resolves multiple data combinations of different dimensions that are orthogonal or weakly correlated. This resolution process removes the influence of the data at the superficial level, accurately decomposing the originally mixed signal into a morphological feature dimension focusing on geometric changes, a texture anomaly dimension focusing on the continuity and fracture of the surface structure, and a stress distribution dimension focusing on the internal physical state of the material. This achieves the physical essence restoration of the detection signal from a "mixed state" to a "decoupled state".

[0037] The system inputs multiple parsed data combinations into a dynamic recognition network, independently executing a multi-scale detection strategy for each dimension of the data combination. During the recognition process, the system constructs a pyramid feature extraction architecture with different receptive fields, performing progressive downsampling and upsampling convolution operations on the data combinations from global to local levels, enabling the dynamic recognition network to simultaneously capture spatial features across scales. Under this dynamic recognition mechanism, the system can identify significant anomalous clusters of large-area energy transitions in the data combinations at the macro scale, and can also keenly capture weak local mutation points hidden in normal background fluctuations at the micro scale. Finally, by fusing multi-scale recognition response maps and spatial coordinate mapping, the system accurately delineates multiple first defect regions corresponding to different physical properties on the silicon wafer surface. These first defect regions refer to the set of suspected anomalous regions initially delineated by the system in the two-dimensional projection surface or initial three-dimensional pose space of the silicon wafer based solely on apparent features such as pixel grayscale mutations, surface texture fractures, and shallow stress optical responses in the multimodal data dimensionality reduction and multi-scale dynamic recognition stages.

[0038] Specifically, for multimodal data of silicon wafers, the high-frequency micro-vibrations of the silicon wafer cause a large amount of non-periodic dynamic displacement noise in the image signals captured by the high-speed camera of the detection device. Direct high-dimensional recognition is prone to misjudgment. Therefore, the system introduces a manifold learning mechanism, such as the temporal neighborhood embedding method, to smoothly reduce the dimensionality of the high-dimensional multimodal data to a three-dimensional intrinsic feature space after removing the random displacement offset caused by vibration. In this intrinsic feature space, the system accurately analyzes the complex signal into a combination of three dimensions based on the physical characteristics of the silicon wafer: the first is the "morphological feature dimension" data combination reflecting the geometric undulations such as edge sagging or center bulging of the silicon wafer; the second is the "texture anomaly dimension" data combination reflecting the micro-groove fractures or foreign matter attachments on the surface of the silicon wafer caused by processing; and the third is the "stress distribution dimension" data combination, which is determined by the difference in birefringence light intensity generated by the multi-angle annular bright and dark field light source matrix and reflects the internal residual stress concentration caused by extreme bending of the silicon wafer.

[0039] The system initiates multi-scale detection for dynamic identification on these three data combinations. Taking the "texture anomaly dimension" and "stress distribution dimension" as examples, at the macro scale, the identification network detects a large area of ​​abnormal stress energy accumulation in the edge region of the silicon wafer with the most severe warping in the "stress distribution dimension" data combination. The system then switches to the micro scale at the corresponding spatial coordinate position and performs a fine scan in the "texture anomaly dimension" data combination. From the strong background scattering light spots caused by warping, it keenly captures several pixel-level weak texture fracture signals, which usually indicate hidden cracks caused by bending stress. The system spatially fuses the macro stress accumulation area and the micro texture fracture point in the image coordinate system to accurately determine and define multiple first defect areas at the edge warping zone of the silicon wafer.

[0040] Furthermore, a corresponding graph neural network is loaded onto multiple first defect regions, and dynamic coupling of multiple first defect regions is triggered, thereby transforming multiple first defect regions into nodes in the topology graph. Feature aggregation is performed on multiple first defect regions simultaneously, and pseudo-defect elimination and boundary fitting are further combined with the constraint relationship of the silicon wafer, thereby outputting the primary defect information of the silicon wafer. The primary defect information of the silicon wafer includes the exact location, shape and physical properties of the defects. The exact location, shape and physical properties of the defects are introduced. At the same time, multimodal data is introduced to control the first defect regions and improve the accuracy of the primary defect information of the silicon wafer.

[0041] At this point, after identifying multiple first defect regions, the system loads a pre-trained graph neural network onto these spatially discrete regions and triggers a dynamic coupling mechanism based on the spatial proximity and feature similarity between regions. During this process, the system abstracts each independent first defect region into an independent node in the topology graph and quantifies the spatial distance and feature correlation between nodes into the weights of the topology edges, thereby constructing a graph structure data that reflects the global distribution state of defects at the logical level. Simultaneously, the graph neural network performs multiple rounds of feature aggregation operations on the edges of the topology graph through a message passing mechanism, so that the feature vector of each node not only contains its own local information, but also deeply integrates the contextual semantics of adjacent nodes, realizing the dimensional evolution from isolated discrete regions to globally correlated features.

[0042] After feature aggregation, the system introduces the inherent physical and geometric constraints of the silicon wafer to filter and correct the nodes in the topology graph. The system uses the crystal orientation symmetry, edge curvature geometry, and the limit deformation threshold allowed by the processing technology of the silicon wafer as hard constraints. By comparing the feature vectors of each aggregated node, the system identifies nodes that violate physical logic or are within the normal texture variation range and judges them as false defects and removes them. For the retained real defect nodes, the system uses its aggregated high-confidence features as a guide and combines the geometric priors provided by the constraints to perform sub-pixel-level boundary fitting and smoothing of the defect edges, eliminating jagged edges and breakpoints in the detection process. Finally, it outputs primary defect information containing the exact spatial location of the defect, the continuous closed shape contour, and the material physical properties obtained by inversion from higher-order features.

[0043] Specifically, for severely warped areas at the edges of silicon wafers, the detection device may identify a dozen or so discrete first defect regions densely on the warped slope due to the complex interference of high-frequency micro-vibrations and multi-angle light sources. At this time, the system loads a graph neural network into these discrete regions, triggering a dynamic coupling mechanism to transform these dozen or so regions into a dozen or so independent nodes in the topological graph. Since real cracks on silicon wafers often appear as discontinuous detection signals under the interference of micro-vibrations, the graph neural network begins to aggregate features through topological edges, logically "bringing together" isolated nodes that are a few millimeters apart but have the same stress direction and texture fracture characteristics, so that the feature vectors of each node incorporate the contextual information of the entire potential anomaly zone.

[0044] The system performs pseudo-defect removal based on the extremely stringent constraints of the silicon wafer: the system calls the silicon wafer... <100> By comparing the isotropic constraints of the crystal orientation and the theoretical threshold of the radius of curvature for edge drooping and warping, it was found that although some nodes in the topology diagram exhibited optical anomalies, their distribution direction completely violated the stress transmission law under the curvature. The system determined that these nodes were pseudo-defects caused by local water stains or light source distortion due to warping, and forcibly removed them from the topology diagram. For the remaining real nodes that conformed to physical laws, the system combined the specific chamfered arc geometric constraints of the silicon wafer edge to perform boundary fitting, smoothing the originally jagged and fragmented node edges caused by vibration into a continuous hidden crack profile that conformed to the chamfered curvature extension. The system accurately output the primary defect information of the silicon wafer: clearly indicating that there is a hidden crack with a definite length and direction at a specific coordinate distance from the edge of the silicon wafer, and marking its physical property as "brittle fracture microcrack caused by stress concentration due to local edge warping".

[0045] In step S13, the specific steps are as follows: S131: Obtain primary defect information, input the primary defect information into the digital twin space along the communication channel, and trigger the time-series deduction mechanism in the digital twin space. The primary defect information is dynamically deduced along the time sequence, and further combined with the material information and mechanical information of the silicon wafer for multi-factor fusion, thereby outputting the defect evolution information of the silicon wafer. The defect evolution information includes the evolution path and risk level. S132: Mark the prediction residual of the defect evolution information, use the prediction residual as a negative feedback signal, the detection device responds to the negative feedback information, and determines the second defect area to be re-inspected in the silicon wafer, and further combines multiple first defect areas to perform point cloud-level registration in three-dimensional space to reconstruct three-dimensional defect features with internal structure mapping relationship.

[0046] In the embodiments of this application, primary defect information is obtained, and the primary defect information is input into the digital twin space along the communication channel. A time-series deduction mechanism is triggered in the digital twin space. The primary defect information is dynamically deduced along the time sequence. Furthermore, the material information and mechanical information of the silicon wafer are combined to perform multi-factor fusion, thereby outputting the defect evolution information of the silicon wafer. The defect evolution information includes the evolution path and risk level, and the evolution path and risk level are introduced.

[0047] At this point, after acquiring the primary defect information containing its precise location, shape, and physical attributes, the system uses a low-latency deterministic communication channel to map and inject the discrete detection results of the physical world into a pre-constructed high-fidelity digital twin space in real time. Within this virtual space, the system accurately locates the virtual entity model that uniquely corresponds to the physical silicon wafer and instantiates the defect features at the corresponding spatial coordinates of the virtual model. This immediately triggers the embedded time-series deduction mechanism, which uses the current detection time as the zero point and discretizes the virtual time axis forward. This drives the virtual defect to perform a step-by-step dynamic deduction along the time sequence in subsequent virtual processing or usage conditions, thereby achieving a leap from static cross-sectional detection to continuous dynamic evolution.

[0048] Driven by time-series simulation, the system synchronously retrieves the microscopic material information and macroscopic mechanical information of the silicon wafer, and introduces these physical parameters as constraint equations into the simulation model for multi-factor fusion calculation. Based on the stress intensity factor and crack propagation theory, the simulation model calculates the stress concentration factor and energy release rate of the defect front at each time step, dynamically simulating the nucleation, subcritical propagation, or unstable fracture process of the defect in virtual spacetime. The system extracts the geometric direction and mechanical attenuation data of the simulation trajectory, and outputs the evolution path of the spatial extension trajectory of the defect in the future time series, as well as the defect evolution information with a quantitative risk level determined by the material failure threshold. Optional microscopic material information of the silicon wafer, such as lattice constant, fracture toughness threshold, and dislocation density distribution; and macroscopic mechanical information, such as elastic modulus, Poisson's ratio, and external load boundary conditions.

[0049] Specifically, when the detection device outputs primary defect information for the silicon wafer, indicating that "a brittle fracture microcrack caused by warping stress exists at a specific coordinate on the edge," this data is instantly injected into the digital twin space via a high-speed optoelectronic communication channel. In the digital twin space, the system "loads this virtual microcrack" at the corresponding edge position of the ultrathin three-dimensional mesh model of the silicon wafer, and triggers a timing simulation mechanism with the current vacuum adsorption state as the time zero point to simulate the state of the silicon wafer in subsequent processes such as wafer transfer and resist removal.

[0050] During the simulation, because the silicon wafer is less than 100 micrometers thick, it is extremely sensitive to local stress. The system integrates the extremely low fracture toughness of the single crystal silicon wafer, the huge bending stress accumulated by the current edge sag and warping, and the local concentrated tensile stress generated when the transfer robot grips it. The simulation model calculates in the virtual time series that at the moment the robot removes the vacuum adsorption, the warping and rebound effect of the silicon wafer edge will cause the microcrack tip to generate an extremely high stress intensity factor, far exceeding the fracture threshold of its extremely thin cross section. Based on this physical calculation, the system outputs the defect evolution information of the silicon wafer: clearly indicating that the evolution path of the microcrack will accelerate its propagation into the silicon wafer along the crystal plane perpendicular to the direction of the maximum principal stress at the edge, and thus determines its risk level as "extremely high risk: failure level".

[0051] Furthermore, the prediction residual of the defect evolution information is marked, and the prediction residual is used as a negative feedback signal. The detection device responds to the negative feedback information and determines the second defect area to be re-inspected in the silicon wafer. Furthermore, multiple first defect areas are combined to perform point cloud-level registration in three-dimensional space to reconstruct three-dimensional defect features with internal structure mapping relationship. This takes into account the overall consideration of negative feedback information and ensures the accuracy of the second defect area to be re-inspected in the silicon wafer.

[0052] At this point, after the defect evolution information is output in the digital twin space, the system calculates the feature deviation between the evolution trajectory and the initial state currently detected by the physical system in real time, thereby quantifying and marking the prediction residual of the defect evolution. The system converts the prediction residual into a negative feedback signal and transmits it downlink. The underlying control architecture of the detection device responds instantly to the negative feedback signal. By analyzing the physical uncertainty represented by the residual, it reversely locates the key spatial coordinates on the silicon wafer where there is missing information or insufficient confidence. The detection device replans the optical observation field of view based on the coordinates and accurately delineates the second defect area to be reviewed on the silicon wafer that needs to be supplemented with feature verification, thereby transforming the deduction blind spot in the virtual space into a targeted review instruction in the physical space. The second defect area refers to the specific area on the silicon wafer that is accurately delineated by the system based on the negative feedback signal during the interaction stage of digital twin temporal deduction and physical space negative feedback. This area is used for in-depth review because a prediction residual is generated between the virtual deduction of the defect evolution trajectory and the actual amount of information contained in the first defect area.

[0053] After identifying the second defect region, the detection device performs high-density feature re-acquisition of this region and projects the acquired supplementary data and the original spatial features of the previously identified multiple first defect regions into a three-dimensional coordinate system. The system performs point cloud-level registration on the depth cues and contour boundaries of all the above discrete regions in three-dimensional space, and eliminates coordinate system rotation and translation errors caused by multi-view observation by using the iterative nearest point method, achieving seamless stitching and spatial fusion of multi-source defect point clouds. Based on the reconstruction of the surface geometric topology, the system further combines the transmittance attenuation and scattering field distribution of the defect under different optical modes to perform voxel-level interpolation into the registered three-dimensional point cloud, and finally reconstructs a three-dimensional defect feature with a mapping relationship from the outer surface morphology to the internal structure.

[0054] Specifically, after the digital twin space deduces the evolution information that the microcracks at the edge of the silicon wafer will accelerate inward, the system marks a huge prediction residual between the "deduced crack propagation tip depth" and the "crack visible depth actually detected by the current detection device". This residual is sent to the detection device as a negative feedback signal instantly. The device recognizes that this negative feedback means that the previous detection failed to see the true bottom of the crack due to the tilt angle obstruction caused by the extreme warping of the silicon wafer. The detection device immediately responds and controls the composite imaging module to perform a greater tilt compensation focus, accurately delineating the specific area obscured below the warped slope, and establishing it as the second defect area to be re-inspected.

[0055] For the second defect region, the detection device switched to a high-penetration specific wavelength light source for intensive scanning, acquiring scattered light spot data from the deep part of the crack. The system then converted this deep data along with the dozen or so first defect regions identified in S121 into three-dimensional spatial coordinates. To address the jitter in point cloud data caused by the high-frequency micro-vibrations unique to silicon wafers, the system adopted a robust point cloud registration method with noise resistance to accurately stitch the deep features of the second defect region with the surface features of multiple first defect regions in three-dimensional space. After eliminating vibration displacement errors, the system used the light and shadow gradient differences at different incident angles provided by the multi-angle annular bright and dark field light source matrix to perform physical depth inversion mapping on the stitched point cloud. Ultimately, the system successfully reconstructed the three-dimensional defect features of the silicon wafer defect, clearly showing not only the forked surface fractures caused by warping stress, but also accurately mapping the three-dimensional spatial structure of the microcracks hidden beneath the warping shadows, which are inverted cone-shaped and penetrate deep into the silicon wafer.

[0056] In step S14, the specific steps are as follows: S141: Collect multiple production processes of silicon wafers, sort the multiple production processes, and simultaneously combine the three-dimensional defect characteristics of silicon wafers to trigger corresponding causal inferences, trace the process chain of silicon wafers in reverse, and thus determine the abnormal processing node that caused the three-dimensional defect. S142: Dynamically trace the abnormal processing node, and combine semantic analysis mechanism to determine multiple abnormal processing information of silicon wafers during the tracing process. Further, through cluster analysis, discover potential common problems in the same batch, and then determine the supplementary detection strategy that the detection device adaptively generates based on the abnormal processing information. The detection device performs supplementary detection on silicon wafers along the supplementary detection strategy. S143: Collect the current working content and corresponding attenuation content of the detection device, determine the current working state of the detection device based on the current working content and corresponding attenuation content, and dynamically plan the global configuration of the detection device for abnormal detection and supplementary detection of the silicon wafer, thereby constructing a closed-loop intelligent detection sequence. This intelligent detection sequence presents an adaptive iterative process between the process fluctuation content of the silicon wafer and the state content of the detection device.

[0057] In the embodiments of this application, multiple production processes of silicon wafers are collected and sorted. Simultaneously, the causal inference triggered by the three-dimensional defect characteristics of the silicon wafer is combined to trace the process chain of the silicon wafer in reverse, thereby determining the abnormal processing node that caused the three-dimensional defect. This approach is compatible with the overall consideration of the causal inference triggered by the three-dimensional defect characteristics of the silicon wafer, ensuring the accuracy of the abnormal processing node that caused the three-dimensional defect.

[0058] At this point, after reconstructing the three-dimensional defect features, the system fully retrieves all production processes that the silicon wafer has undergone throughout the entire processing cycle through the data bus of the manufacturing execution system. It then strictly sorts these discrete process nodes according to timestamps and process logic dependencies, constructing a complete process link with temporal constraints. Simultaneously, the system uses the spatial morphology, depth direction, and mechanical properties in the reconstructed three-dimensional defect features as the logical starting point to trigger the embedded causal inference engine. It uses the physical appearance of the three-dimensional defect as the result variable and prepares to solve for the source cause of the result in the sorted process link, thereby initiating a reverse tracing process from result to cause.

[0059] Driven by causal inference, the system traces back along the sorted process links in reverse time, cross-comparing and counterfactually reasoning the multidimensional parameters of the three-dimensional defect features with the standard process physical models of each process node. The system focuses on analyzing whether the mechanical force field, thermal stress field, or chemical removal rate of a specific process has sufficient conditions to produce the corresponding morphology of the three-dimensional defect. By calculating the probability of causal correlation between each process and the defect features, normal process nodes that contribute very little to the formation of the defect are removed. After reversely penetrating multiple layers of process links, the system identifies specific process nodes whose causal correlation probability exceeds a preset threshold and accurately identifies them as abnormal processing nodes that cause the three-dimensional defect.

[0060] Specifically, after the detection device accurately reconstructs the three-dimensional defect characteristics of the silicon wafer, namely "the surface is bifurcated and fragmented, and the interior is inverted conical and penetrates deep into the interior", the system immediately collects all the production processes that the silicon wafer has undergone from the upstream MES system, and strictly sorts them into a rigorous topological sequence according to the process timeline of "single crystal pulling > slicing > chamfering > rough grinding > fine grinding > edge polishing > double-sided thinning > final cleaning", forming a process link with a clear timeline; simultaneously, the system uses the "depth penetration angle" and "surface stress release morphology" of the inverted conical three-dimensional defect as trigger conditions to start the causal inference engine for reverse tracing.

[0061] The causal inference engine ruled out the "final cleaning" and "edge polishing" processes, which do not involve mechanical intrusion. Continuing back to the "double-sided thinning" process, the system introduced counterfactual reasoning by considering the extremely thin characteristics of silicon wafers (less than 100 micrometers). It determined that while simple in-plane thinning stress could cause overall warping, it could not form a clearly directional inverted conical internal microcrack. When tracing back to the "rough grinding" process node, the system extracted the spatial orientation angle of the three-dimensional inverted conical defect and performed a spatial geometric intersection with the cutting trajectory of the grinding wheel in the rough grinding process. The causal inference model calculated that the longitudinal depth of the inverted conical microcrack highly coincided with the distribution of the micro-cutting stress field generated by a grinding wheel of a specific grit size in the rough grinding stage, and the surface bifurcation and fragmentation perfectly matched the physical logic of releasing the latent damage from the early grinding process in subsequent thinning steps. Based on this, the system, with a very high probability of causal correlation, penetrated multiple process links and accurately determined that the abnormal processing node causing the three-dimensional defect in the silicon wafer was a "specific local over-cutting event in the rough grinding process."

[0062] Furthermore, the abnormal processing node is dynamically traced, and during the tracing process, a semantic analysis mechanism is used to determine multiple abnormal processing information of the silicon wafer. Then, cluster analysis is used to uncover potential common problems in the same batch, and a supplementary detection strategy is determined based on the abnormal processing information adaptively generated by the detection device. The detection device performs supplementary detection on the silicon wafer along the supplementary detection strategy, which takes into account the overall consideration of the abnormal processing node and ensures the accuracy of multiple abnormal processing information of the silicon wafer.

[0063] At this point, the system checks the equipment history database and sensor log network at the bottom of the workshop to perform high-precision dynamic tracing of the processing time window of the abnormal processing node. During this tracing process, the system introduces a semantic analysis mechanism from the field of natural language processing, jointly constructing a semantic knowledge graph from unstructured equipment alarm text, operator maintenance notes, and structured sensor time-series parameters. By extracting entity words from the text logs and performing semantic alignment and joint disambiguation with the abnormal fluctuation features in the structured parameters, the system accurately parses out the specific physical causes and environmental context that led to the abnormality of the node, thereby outputting abnormal processing information in a structured manner that includes multiple dimensions such as equipment state deviation, consumable failure characteristics, or process parameter exceeding limits. Entity words in the text logs include equipment component names and error codes.

[0064] After extracting multidimensional abnormal processing information, the system further calls upon cross-batch production big data. Using the feature vector of the abnormal processing information as the cluster center, density clustering is used to search for sample sets with similar feature spatial distances in the historical database, thereby uncovering common problems hidden in the same batch or under machines using the same consumables. Based on the scope of this common problem and the evolution law of defects, the system combines the current optical module margin and kinematic constraints of the detection device to adaptively generate targeted supplementary detection strategies. This strategy dynamically reconstructs the scanning path, light source illumination combination, and sampling resolution parameters of the detection device to form a new sequence of control commands, driving the detection device to deviate from the default detection process and strictly follow the supplementary detection strategy to perform physical-level supplementary detection on the current silicon wafer or related silicon wafers.

[0065] Specifically, after determining that the abnormal processing node causing the inverted conical microcrack inside the silicon wafer was the "rough grinding process", the system immediately dynamically traced the machine's underlying data at the time of execution of this process. Since the grinding machine log contained both structured speed and current fluctuation curves and unstructured "amplitude over-limit shutdown and reset" alarm text, the system activated a semantic analysis mechanism. This mechanism semantically fused and aligned the semantic entity "amplitude over-limit" in the text with the abnormal high-frequency oscillation features of a specific frequency band in the current curve, accurately parsing out the specific abnormal processing information: that is, a certain bonding pad of the rough grinding machine had a "transient high-frequency resonance impact caused by the local solidified abrasive falling off" at a specific angle.

[0066] The system uses the feature vector of the "high-frequency resonance impact at a specific angle" as its core to perform density clustering analysis on all silicon wafer production records that passed through the coarse grinding machine in that shift. This analysis uncovers a potential common problem: the resonance causes a very high risk of microcracks appearing at the corresponding angle in silicon wafers of the same batch and in a specific crystal orientation. Based on this common hidden danger, the system adaptively generates a supplementary detection strategy, taking into account the current hardware capabilities of the detection device. Since conventional detection may miss similar microcracks in non-warped areas, the strategy requires the detection device's multi-angle annular bright-dark field light source matrix to forcibly switch to "high-frequency flicker dark field mode," and instructs the camera with tilt compensation function to perform a wafer-by-wafer encrypted scan of the "specific resonance angle region" of that batch of silicon wafers. Upon receiving this strategy, the detection device immediately interrupts the conventional production line operation and performs highly sensitive supplementary detection on other areas of the current silicon wafer along the aforementioned special lighting and tilt tracking path to thoroughly check for any missed defects caused by the same resonance problem.

[0067] Therefore, the current working content and corresponding attenuation content of the detection device are collected. Based on the current working content and corresponding attenuation content of the detection device, the current working state of the detection device is determined, and the global configuration of the detection device for abnormal detection and supplementary detection of the silicon wafer is dynamically planned. This constructs a closed-loop intelligent detection sequence. This intelligent detection sequence presents an adaptive iterative process between the process fluctuation content of the silicon wafer and the state content of the detection device. It is compatible with the overall consideration of the current working content and corresponding attenuation content of the detection device, ensuring the accuracy of the current working state of the detection device. At the same time, it further controls the second defect area, fully considers multiple abnormal processing information and the current working state of the detection device, and improves the accuracy of the intelligent detection sequence.

[0068] At this moment, while triggering the supplementary detection command, the system collects the current working content accumulated by the detection device during the execution of the preceding anomaly detection and supplementary detection in real time. This content quantitatively represents the density of the mechanism's motion trajectory and the excitation frequency of the optical module within a specific time period. Simultaneously, the system extracts the physical feedback from the device's underlying sensors and calculates the attenuation content bound to the aforementioned working content. This attenuation content specifically covers the luminous flux attenuation rate of the light source system, the positioning accuracy drift of the motion axis system, and the thermal noise accumulation level of the image sensor. The system inputs the high-load working content and multi-dimensional attenuation content into the state evaluation model. By calculating the degree to which the device performance deviates from the ideal threshold, the system accurately derives and determines the current real working state of the detection device that can support high-quality imaging at this moment.

[0069] The system abandons static, fixed detection logic and uses the current attenuation limit as a hard constraint boundary to dynamically plan the global configuration of subsequent anomaly detection and supplementary detection tasks for the silicon wafer. This planning process performs multi-objective optimization between the silicon wafer's process fluctuations and the state of the detection device, adaptively reallocating detection resources, such as dynamically reducing the sampling rate of non-critical areas to compensate for the light source energy of critical areas. The system stitches together the preceding anomaly identification, twin deduction, root cause tracing, supplementary detection, and the current resource reallocation based on the device state on the temporal and logical axes to construct a closed-loop intelligent detection sequence. This sequence physically presents the entire adaptive iterative process of how silicon wafer process fluctuations drive detection strategy adjustments and how the attenuation of the detection device state inversely corrects the detection threshold. The silicon wafer's process fluctuations refer to the severity of defect evolution and the urgency of supplementary detection; the detection device's state refers to the remaining effective resolution and focus response bandwidth.

[0070] Specifically, when the detection device performs "specific angle encrypted supplementary detection" on the silicon wafer caused by coarse grinding resonance, the system collects the current working content of the composite shooting module in real time during this period: that is, the tilt servo motor performs extremely high frequency reciprocating deflection to track the warped surface, and the ring dark field light source is in a continuous high-energy pulse excitation state; based on this, the system simultaneously extracts the corresponding attenuation content: it is found that due to long-term high-frequency shooting, the piezoelectric ceramic of the Z-axis fast response focusing mechanism has micron-level thermal expansion positioning drift.

[0071] The system incorporates the attenuation factors of "high-load reciprocating deflection" and "positioning drift" into the state assessment model and determines that the current physical working state of the detection device can no longer support the original ultra-high precision microcrack identification task. At this point, based on the high risk brought about by silicon wafer process fluctuations and the existence of batch defects caused by resonance, the system performs dynamic planning of the global configuration of the subsequent detection path: the system decides not to perform uniform high-precision scanning on the entire surface of the silicon wafer, but to actively avoid the equipment attenuation bottleneck and adaptively concentrate all remaining detection resources on the highest risk area of ​​"rough grinding resonance angle". At the same time, in response to focus drift and light decay, the system dynamically lowers the edge contrast extraction threshold in the global configuration and instructs the tilt camera to determine a single optimal vibration avoidance angle to reduce mechanical heat accumulation.

[0072] The system seamlessly integrates the targeted supplementary detection in step S142 with the resource reallocation actions such as "frequency reduction, determination, and threshold adjustment" performed in step S143 based on light source attenuation and mechanism drift, thus completely constructing a closed-loop intelligent detection sequence. In this sequence, the extremely high "process fluctuation content" caused by the ultra-thin warp and grinding resonance of the silicon wafer successfully induces the detection device to overdraw its performance for supplementary detection. The "light decay and drift state content" generated by the detection device during execution serves as a negative constraint, forcing the system to modify the subsequent detection parameters for the silicon wafer, perfectly realizing the deep adaptive iteration and dynamic balance between silicon wafer process fluctuations and equipment physical state.

[0073] Please see Figure 2 The intelligent detection system of the detection device is applied to the intelligent detection method of the above-mentioned detection device; the intelligent detection system of the detection device includes: The multimodal data module 21 is used to adjust the composite imaging module of the detection device according to the posture data of the silicon wafer during the detection process of the detection device, synchronously acquire real-time images of the silicon wafer, and determine multimodal data by combining the production process of the silicon wafer and the corresponding surface texture. The primary defect information module 22 is used to determine multiple data combinations of different dimensions based on the parsing of the multimodal data, determine the corresponding first defect region based on the identification of each data combination, dynamically couple multiple first defect regions, and further combine the constraint relationship of the silicon wafer to determine the primary defect information of the silicon wafer. The defect region module 23 is used to input primary defect information into the digital twin space and perform dynamic simulation in combination with the time-series simulation mechanism, thereby outputting the defect evolution information of the silicon wafer. Based on the negative feedback of the defect evolution information, the second defect region to be re-examined is determined, and the corresponding three-dimensional defect features are further determined by combining multiple first defect regions. The intelligent detection module 24 is used to mark multiple production processes of silicon wafers, determine the corresponding abnormal processing nodes by combining the three-dimensional defect characteristics of silicon wafers, determine multiple abnormal processing information corresponding to silicon wafers by tracing the abnormal processing nodes, further trigger supplementary detection of the detection device, and construct a closed-loop intelligent detection sequence by combining the current working status of the detection device.

[0074] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent detection method for a detection device, characterized in that, include: During the inspection of silicon wafers by the inspection device, the composite imaging module of the inspection device is adjusted according to the orientation data of the silicon wafer to synchronously acquire real-time images of the silicon wafer, and multimodal data is determined by combining the production process of the silicon wafer and the corresponding surface texture. Based on the analysis of the multimodal data, multiple data combinations of different dimensions are determined. Based on the identification of each data combination, the corresponding first defect region is determined. Multiple first defect regions are dynamically coupled, and the primary defect information of the silicon wafer is further determined by combining the constraint relationship of the silicon wafer. The primary defect information is input into the digital twin space and dynamically extrapolated using a time-series extrapolation mechanism, thereby outputting the defect evolution information of the silicon wafer. Based on the negative feedback of this defect evolution information, the second defect area to be re-examined is determined, and the corresponding three-dimensional defect features are further determined by combining multiple first defect areas. Multiple production processes of silicon wafers are marked, and the corresponding abnormal processing nodes are determined by combining the three-dimensional defect characteristics of the silicon wafers. Based on the tracing of the abnormal processing nodes, multiple abnormal processing information corresponding to the silicon wafers are determined, which further triggers supplementary detection by the detection device. A closed-loop intelligent detection sequence is constructed by combining the current working status of the detection device.

2. The intelligent detection method of the detection device according to claim 1, characterized in that, During the inspection of the silicon wafer by the inspection device, the composite imaging module of the inspection device is adjusted according to the orientation data of the silicon wafer to synchronously acquire real-time images of the silicon wafer, and multimodal data is determined by combining the production process of the silicon wafer and the corresponding surface texture, including: When the inspection device performs silicon wafer inspection, it acquires the three-dimensional posture data of the silicon wafer at the inspection station in real time based on the posture sensor. Multiple inspection directions are determined based on the recognition of the three-dimensional posture data of the silicon wafer. The three-dimensional posture data of the silicon wafer is used as a feedforward signal to trigger the control of the composite imaging module of the inspection device, so that the composite imaging module of the inspection device can adaptively inspect the silicon wafer along different angles to acquire real-time images of the silicon wafer.

3. The intelligent detection method of the detection device according to claim 2, characterized in that, During the inspection of silicon wafers by the inspection device, the composite imaging module of the inspection device is adjusted according to the orientation data of the silicon wafer to synchronously acquire real-time images of the silicon wafer, and multimodal data is determined by combining the production process of the silicon wafer and the corresponding surface texture. This also includes: The process involves acquiring the production history and surface texture of silicon wafers, aligning the surface texture with real-time images of the wafers in time and space, dynamically constructing a corresponding data framework, and further integrating multiple factors with the production history of the silicon wafers to build multimodal data with high-order correlation.

4. The intelligent detection method of the detection device according to claim 1, characterized in that, The process involves determining multiple data combinations of different dimensions based on the analysis of the multimodal data, identifying corresponding first defect regions based on the identification of each data combination, dynamically coupling multiple first defect regions, and further combining the constraints of the silicon wafer to determine the primary defect information of the silicon wafer, including: A manifold learning mechanism was introduced for the multimodal data, which was then dimensionality-reduced and mapped to simultaneously resolve multiple data combinations of different dimensions. These multiple data combinations covered the dimensions of morphological features, texture anomalies, and stress distribution. Each data combination was dynamically identified, and multi-scale detection was performed during the identification process to determine multiple first defect regions of the silicon wafer.

5. The intelligent detection method of the detection device according to claim 4, characterized in that, The process of determining multiple data combinations of different dimensions based on the analysis of the multimodal data, determining the corresponding first defect region based on the identification of each data combination, dynamically coupling multiple first defect regions, and further determining the primary defect information of the silicon wafer by combining the constraint relationship of the silicon wafer, also includes: A graph neural network is loaded for multiple first defect regions, and dynamic coupling of multiple first defect regions is triggered, thereby transforming multiple first defect regions into nodes in the topology graph. Feature aggregation is performed on multiple first defect regions simultaneously, and pseudo-defect elimination and boundary fitting are further combined with the constraint relationship of the silicon wafer, thereby outputting the primary defect information of the silicon wafer. The primary defect information of the silicon wafer includes the exact location, shape and physical properties of the defects.

6. The intelligent detection method of the detection device according to claim 1, characterized in that, The process involves inputting primary defect information into a digital twin space and dynamically extrapolating it using a time-series deduction mechanism to output defect evolution information for the silicon wafer. Based on the negative feedback of this defect evolution information, a second defect region to be re-inspected is determined. Furthermore, multiple first defect regions are combined to determine the corresponding three-dimensional defect features, including: Primary defect information is acquired and input into the digital twin space through the communication channel. A time-series deduction mechanism is triggered in the digital twin space, and the primary defect information is dynamically deduced along the time sequence. Furthermore, the material and mechanical information of the silicon wafer is combined to perform multi-factor fusion, thereby outputting the defect evolution information of the silicon wafer. This defect evolution information includes the evolution path and risk level.

7. The intelligent detection method of the detection device according to claim 6, characterized in that, The process of inputting primary defect information into a digital twin space and performing dynamic simulation using a time-series extrapolation mechanism to output defect evolution information of the silicon wafer, determining the second defect region to be re-inspected based on the negative feedback of this defect evolution information, and further determining the corresponding three-dimensional defect features by combining multiple first defect regions, also includes: The prediction residual, which is labeled with the defect evolution information, is used as a negative feedback signal. The detection device responds to the negative feedback information and determines the second defect region to be re-examined in the silicon wafer. Furthermore, it combines multiple first defect regions to perform point cloud-level registration in three-dimensional space to reconstruct three-dimensional defect features with internal structure mapping relationships.

8. The intelligent detection method of the detection device according to claim 1, characterized in that, The multiple production processes of the marked silicon wafer, combined with the three-dimensional defect characteristics of the silicon wafer, determine the corresponding abnormal processing nodes. Based on the tracing of these abnormal processing nodes, multiple abnormal processing information corresponding to the silicon wafer are determined, further triggering supplementary detection by the detection device. A closed-loop intelligent detection sequence is constructed based on the current operating status of the detection device, including: Collect data from multiple production processes of silicon wafers, sort these processes, and simultaneously combine them with the three-dimensional defect characteristics of the silicon wafers to trigger corresponding causal inferences. Then, trace the process chain of the silicon wafers in reverse to identify the abnormal processing nodes that caused the three-dimensional defect. The abnormal processing node is dynamically traced, and during the tracing process, a semantic analysis mechanism is used to determine multiple abnormal processing information of the silicon wafer. Furthermore, cluster analysis is used to uncover potential common problems in the same batch, and then a supplementary detection strategy is determined based on the abnormal processing information that the detection device adaptively generates. The detection device then performs supplementary detection on the silicon wafer according to the supplementary detection strategy.

9. The intelligent detection method of the detection device according to claim 8, characterized in that, The process of marking multiple production steps of the silicon wafer, combined with the three-dimensional defect characteristics of the silicon wafer, determines the corresponding abnormal processing nodes. Based on the tracing of these abnormal processing nodes, multiple abnormal processing information corresponding to the silicon wafer are determined, further triggering supplementary detection by the detection device. A closed-loop intelligent detection sequence is constructed based on the current operating state of the detection device. The process also includes: The system collects the current working content and corresponding attenuation content of the detection device, determines the current working state of the detection device based on the current working content and corresponding attenuation content, and dynamically plans the global configuration of the detection device for abnormal detection and supplementary detection of the silicon wafer, thereby constructing a closed-loop intelligent detection sequence. This intelligent detection sequence presents an adaptive iterative process between the process fluctuation content of the silicon wafer and the state content of the detection device.

10. An intelligent detection system for a detection device, characterized in that, The intelligent detection system of the detection device is applied to the intelligent detection method of the detection device as described in any one of claims 1-9; The intelligent detection system of the detection device includes: The multimodal data module is used to adjust the composite imaging module of the inspection device according to the posture data of the silicon wafer during the inspection process of the inspection device, synchronously acquire real-time images of the silicon wafer, and determine multimodal data by combining the production process of the silicon wafer and the corresponding surface texture. The primary defect information module is used to determine multiple data combinations of different dimensions based on the parsing of the multimodal data, determine the corresponding first defect region based on the identification of each data combination, dynamically couple multiple first defect regions, and further combine the constraint relationship of the silicon wafer to determine the primary defect information of the silicon wafer. The defect region module is used to input primary defect information into the digital twin space and perform dynamic simulation in combination with the time-series simulation mechanism, thereby outputting the defect evolution information of the silicon wafer. Based on the negative feedback of the defect evolution information, the second defect region to be re-examined is determined, and the corresponding three-dimensional defect features are further determined by combining multiple first defect regions. The intelligent detection module is used to mark multiple production processes of silicon wafers, determine the corresponding abnormal processing nodes by combining the three-dimensional defect characteristics of silicon wafers, determine multiple abnormal processing information corresponding to silicon wafers by tracing the abnormal processing nodes, further trigger supplementary detection by the detection device, and construct a closed-loop intelligent detection sequence by combining the current working status of the detection device.

Images