Screening device, screening method based on multiple working conditions and system

Abstract

The application discloses a kind of screening equipment based on the screening method and system of multiple working conditions, and the application relates to the technical field of multiple working conditions processing, the current state of screening equipment is determined according to the multi-level iteration of the digital twin body;According to the analysis of digital twin body, the image content of each silicon wafer is determined, the global feature spectrum of the silicon wafer is determined along the dynamic identification of the image content, and the screening condition of the silicon wafer by the screening equipment is further determined by combining the convolutional neural network preset by the screening equipment, which improves the accuracy of the screening condition of the silicon wafer by the screening equipment. Mark the screening condition after adjustment, determine the multiple online working parameters of the screening equipment according to the traceability of the screening condition, and combine the corresponding screening result of the real-time output of the multi-source heterogeneous visual array, determine the corresponding screening deviation information according to the comparison of the screening result in different dimensions, and combine the corresponding multi-working condition screening system by building block chain, which improves the accuracy of the multi-working condition screening system.

Description

Technical Field

[0001] This invention relates to the technical field of multi-condition processing, and in particular to a screening method and system based on multi-condition screening equipment. Background Technology

[0002] As a fundamental material in the photovoltaic industry, the production quality of silicon wafers directly determines the photoelectric conversion efficiency of photovoltaic modules. In the silicon wafer production process, the sieving process is a key step to ensure that silicon wafers are graded according to thickness, size, and appearance quality. With the increasingly stringent requirements of the photovoltaic industry for silicon wafer quality, traditional sieving equipment and control methods are facing many challenges.

[0003] Most existing technologies employ a single visual recognition method, lacking in-depth perception and integration of the operating status of the screening equipment itself. In actual production, screening equipment is often under complex conditions such as high speed and vibration. The vibration frequency of the equipment itself, the fluctuation of the conveyor belt speed, and the distribution of mechanical stress directly affect the dynamic imaging quality of the silicon wafer. The failure to construct a digital twin that includes the physical constraints of the entire "optical-mechanical-electrical" chain affects the accuracy of the screening equipment in screening silicon wafers, leading to the accuracy of the multi-condition screening system. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a screening method and system based on multiple working conditions for screening equipment.

[0005] This invention provides a screening method for a screening device based on multiple operating conditions, including: Collect multiple working data of the screening equipment at the current moment, and combine them with multiple image data of the screening equipment at the screening station to create a digital twin, so as to construct a corresponding digital twin; determine the current state of the screening equipment based on the multi-level iteration of the digital twin; Based on the analysis of the digital twin, the image content of each silicon wafer is determined. The global feature map of the silicon wafer is determined by the dynamic recognition of the image content. Furthermore, the screening conditions of the screening equipment for silicon wafers are determined by the pre-set convolutional neural network of the screening equipment. The screening conditions include fine screening, general screening, and waste removal. The screening conditions of the silicon wafers and the current state of the screening equipment are input into a unified control space, and a fuzzy PID control mechanism is triggered in the control space. When risk information of silicon wafers is detected, the screening conditions of the screening equipment are adaptively adjusted, thereby realizing the adaptive mapping between silicon wafers of different product grades and the operating parameters of the screening equipment. The adjusted screening conditions are marked, and multiple online operating parameters of the screening equipment are determined based on the traceability of the screening conditions. The corresponding screening results are output in real time by combining the multi-source heterogeneous vision array. The corresponding screening deviation information is determined by comparing the screening results in different dimensions, and the corresponding multi-condition screening system is constructed by combining blockchain.

[0006] This invention provides a screening system based on multiple operating conditions, which is applied to the aforementioned screening method based on multiple operating conditions. The screening system includes: The digital twin module is used to collect multiple working data of the screening equipment at the current moment, and combine them with multiple image data of the screening equipment at the screening station to create a digital twin to construct a corresponding digital twin; the current state of the screening equipment is determined based on the multi-level iteration of the digital twin. The working condition module is used to determine the image content of each silicon wafer based on the analysis of the digital twin, determine the global feature map of the silicon wafer by dynamically recognizing the image content, and further combine the pre-set convolutional neural network of the screening equipment to determine the screening working condition of the screening equipment for the silicon wafer. The screening working condition covers fine screening working condition, general screening working condition, and waste removal working condition. The adaptive adjustment module is used to input the screening conditions of the silicon wafers and the current state of the screening equipment into a unified control space, and trigger a fuzzy PID control mechanism in the control space. This triggers the adaptive adjustment of the screening conditions of the silicon wafers when risk information of the silicon wafers is detected, thereby realizing the adaptive mapping between the operating parameters of silicon wafers of different product grades and the screening equipment. The multi-condition screening system module is used to mark the adjusted screening conditions, determine multiple online working parameters of the screening equipment based on the traceability of the screening conditions, and output the corresponding screening results in real time in combination with a multi-source heterogeneous vision array. Based on the comparison of the screening results in different dimensions, the corresponding screening deviation information is determined, and the corresponding multi-condition screening system is constructed in combination with blockchain.

[0007] Compared with the prior art, the beneficial effects of the present invention are: (1) Collect multiple working data of the screening equipment at the current moment, and combine them with multiple image data of the screening equipment at the screening station to form a digital twin to construct a corresponding digital twin; determine the current state of the screening equipment based on the multi-level iteration of the digital twin; determine the image content of each silicon wafer based on the analysis of the digital twin, determine the global feature map of the silicon wafer by following the dynamic recognition of the image content, and further combine the pre-set convolutional neural network of the screening equipment to determine the screening condition of the silicon wafer. The introduction of the digital twin further controls the global feature map of the silicon wafer and improves the accuracy of the screening condition of the silicon wafer.

[0008] (2) The screening conditions of the silicon wafers and the current status of the screening equipment are input into a unified control space, and a fuzzy PID control mechanism is triggered in the control space. When the risk information of the silicon wafer is detected, the screening equipment is triggered to adaptively adjust the screening conditions of the silicon wafers. The adjusted screening conditions are marked. Multiple online working parameters of the screening equipment are determined based on the traceability of the screening conditions. The corresponding screening results are output in real time by combining the multi-source heterogeneous vision array. The corresponding screening deviation information is determined by comparing the screening results in different dimensions. The corresponding multi-condition screening system is constructed by combining blockchain. The adaptive mapping between silicon wafers of different product grades and the operating parameters of the screening equipment is realized. The screening deviation information and blockchain are fully considered, which improves the accuracy of the multi-condition screening system. Attached Figure Description

[0009] Figure 1 This is a schematic flowchart of a screening method based on multiple working conditions in a screening device according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structural composition of the screening system based on multiple working conditions in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Please see Figures 1 to 2 A screening device based on a multi-condition screening method is applied to multi-condition processing scenarios; the multi-condition screening device includes: Step S11: Collect multiple working data of the screening equipment at the current moment, and combine them with multiple image data of the screening equipment at the screening station to create a digital twin, so as to construct a corresponding digital twin; determine the current state of the screening equipment based on the multi-level iteration of the digital twin; Step S12: Determine the image content of each silicon wafer based on the analysis of the digital twin, determine the global feature map of the silicon wafer by dynamically recognizing the image content, and further combine the pre-set convolutional neural network of the screening equipment to determine the screening conditions of the screening equipment for the silicon wafers. The screening conditions include fine screening, general screening, and waste removal. Step S13: Input the screening conditions of the silicon wafers by the screening equipment and the current status of the screening equipment into a unified control space, and trigger the fuzzy PID control mechanism in the control space. This triggers the adaptive adjustment of the screening conditions of the silicon wafers by the screening equipment when the risk information of the silicon wafers is detected, thereby realizing the adaptive mapping between the operating parameters of silicon wafers of different product grades and the screening equipment. Step S14: Mark the adjusted screening conditions, determine multiple online working parameters of the screening equipment based on the traceability of the screening conditions, and output the corresponding screening results in real time in combination with the multi-source heterogeneous vision array. Determine the corresponding screening deviation information based on the comparison of the screening results in different dimensions, and build the corresponding multi-condition screening system in combination with blockchain.

[0012] In step S11, the specific steps are as follows: S111: Real-time monitoring of the screening equipment, and determination of multiple working data of the screening equipment at the current moment based on the detection of the screening equipment database. The multiple working data cover vibration frequency, conveyor belt speed, execution action and dynamic detection data; marking the multi-source heterogeneous vision array of the screening equipment, and dynamically capturing images of silicon wafers in the screening area based on the multi-source heterogeneous vision array, presenting multiple image data of silicon wafers under high-speed movement, and presenting an image stream of silicon wafers at different angles; S112: Spatiotemporal alignment of multiple working data and multiple image data is performed, and a digital twin mechanism is introduced in the alignment process to dynamically construct a digital twin with fluid-solid coupling and optical-mechanical-electrical full-link physical constraints; S113: The digital twin undergoes multi-level iterations, successively going through primary iteration, intermediate iteration, and advanced iteration. At this point, the primary iteration performs spatiotemporal registration and noise filtering to restore the real scene. The intermediate iteration uses implicit neural technology to fill in image blind spots and deduce the invisible stress distribution inside the equipment. The advanced iteration introduces a long short-term memory mechanism to predict the future microsecond-level evolution of the equipment state based on the trajectory of previous operating conditions, thereby determining the current state of the screening equipment. The current state of the screening equipment not only includes the current working state of the screening equipment, but also embeds a tensor expression of the execution deviation.

[0013] In the embodiments of this application, the screening equipment is monitored in real time, and multiple working data of the screening equipment at the current moment are determined based on the detection of the screening equipment's database. These multiple working data cover vibration frequency, conveyor belt speed, execution actions, and dynamic detection data. The multi-source heterogeneous vision array of the screening equipment is marked, and the silicon wafers in the screening area are dynamically photographed based on the multi-source heterogeneous vision array. Multiple image data of the silicon wafers under high-speed movement are presented, showing the image flow of the silicon wafers at different angles. This is compatible with the overall consideration of the screening equipment's database and ensures the accuracy of the multiple working data of the screening equipment at the current moment.

[0014] At this time, the system establishes a real-time data link through sensor groups deployed at each key node of the screening equipment; the system polls the field controller and historical database of the screening equipment at a high frequency sampling rate, and locks the current moment's equipment operating status snapshot through a timestamp synchronization mechanism; the sensor groups at each key node include vibration sensors, photoelectric encoders, current transformers, etc.

[0015] Meanwhile, vibration frequency data is obtained through a triaxial accelerometer mounted on the vibrating motor base of the screening unit, reflecting the excitation intensity and stability of the equipment; conveyor belt speed data is fed back in real time through the photoelectric encoder of the conveyor belt drive motor, used to calculate the dynamic displacement of the silicon wafers; execution action data comes from the control signal feedback of the pneumatic sorting valve or the robotic arm, recording the instantaneous action logic of the screening mechanism; dynamic detection data covers process quantities such as motor load current and bearing temperature, which together constitute the "physical state vector" of the equipment.

[0016] The system performs logical marking and calibration on a multi-source heterogeneous vision array deployed above and to the side of the screening station. This array includes a high-speed linear scan camera and an area scan camera. For silicon wafers in a high-speed transmission state within the screening area, the vision array triggers synchronous exposure control to overcome motion blur and deformation caused by the high-speed movement of the silicon wafer. Through microsecond-level synchronization of the exposure time of the stroboscopic light source and the camera, the system acquires a continuous sequence of images of the silicon wafer from multiple perspectives on the front, side, and back. These image data are not isolated single-frame images, but rather an image stream that strictly conforms to temporal logic, fully presenting the edge contours, surface microcracks, and chamfer features of the silicon wafer under high-speed movement. The high-speed linear scan camera is used to capture continuous motion images, while the area scan camera is used to acquire local high-frame-rate details.

[0017] Specifically, during the operation of the screening equipment, the system performs work data extraction. Assuming the screening equipment is currently operating in "fine screening mode" and the conveyor belt speed is set to 1.5 m / s, the system reads the database and sensor data in real time, confirming that the vibration frequency is stable at 50 Hz and the conveyor belt speed feedback value is 1.498 m / s, which is within the allowable error range. The execution action signal shows that the sorting air valves at each level are in standby mode, and the dynamic detection data shows that the main drive motor current is stable and there are no signs of overload. These data confirm that the equipment is at a healthy operating baseline.

[0018] Meanwhile, the multi-source heterogeneous vision array begins to operate; when a batch of silicon wafers enters the sorting area at high speed, the front high-speed linear array camera and the side area array camera in the vision array work together; for a silicon wafer traveling at a speed of 1.5 meters per second, the camera captures images at a shutter speed in the microsecond range, obtaining image streams of the silicon wafer at different angles; the front image stream clearly shows the hidden cracks and texture features on the surface of the silicon wafer, while the side image stream captures the chipping at the edge of the silicon wafer. These image stream data and the vibration and belt speed data acquired at the front end correspond strictly to each other on the time axis, jointly constructing the original digital base of silicon wafer sorting at that moment, providing an accurate physical world mapping for the subsequent construction of digital twins.

[0019] Furthermore, multiple working data and multiple image data are spatiotemporally aligned, and a digital twin mechanism is introduced during the alignment process to dynamically construct a digital twin with fluid-solid coupling and optical-mechanical-electrical full-link physical constraints.

[0020] At this point, the system establishes a unified spatiotemporal coordinate system to solve the data misalignment problem caused by inconsistent sensor sampling frequencies and transmission delays. The system introduces a high-precision clock synchronization protocol to map low-frequency one-dimensional time-series data such as vibration frequency and conveyor belt speed to the same time axis with high-frequency two-dimensional image stream data. At the same time, the kinematic model of the screening equipment is used to transform the data deviation in the time dimension into position compensation in the spatial dimension. For example, based on the time difference between the conveyor belt speed and the image acquisition time, the actual lag distance of the silicon wafer in physical space is calculated, thereby correcting the spatial coordinates of the image data in virtual space. This ensures that the "behavioral data" of the equipment and the "visual data" of the silicon wafer are fully coupled at the millisecond level, forming a spatiotemporally aligned fusion dataset.

[0021] After data alignment is completed, the system dynamically constructs a virtual entity based on a digital twin engine. This entity not only has a geometric appearance but also embeds multi-physics coupling equations. The system introduces full-link physical constraints of opto-mechanical-electronic, that is, it establishes the transfer function relationship between motor drive voltage, vibration frequency and light response in the virtual model to ensure that the drive response characteristics of the virtual device are consistent with the physical entity. At the same time, for the airflow sorting link in the screening process, the system injects a fluid-solid coupling mechanism, uses computational fluid dynamics to simulate the interaction force between the screening airflow field and the silicon wafer solid rigid body, and reconstructs the perturbation model of airflow on silicon wafer attitude in the virtual environment. In this way, a high-fidelity digital twin that conforms to both geometric similarity and physical laws is constructed.

[0022] Specifically, when a target silicon wafer passes through the screening area at high speed on the conveyor belt, the system performs a spatiotemporal alignment operation. Assuming that the image acquired by the vision array shows that there is a micron-level chip at the edge of the silicon wafer, and the data fed back by the vibration sensor at this moment has a transmission delay of about 5 milliseconds, the system uses the spatiotemporal alignment method to calculate the spatial displacement corresponding to these 5 milliseconds as 7.5 millimeters based on the conveyor belt speed. In this way, the image feature is backtracked and corrected in digital space to ensure that the "silicon wafer in the image" and the "silicon wafer on the vibration table" are strictly coincident in spatial position.

[0023] The system dynamically constructs a digital twin. In the virtual environment, the system activates the opto-mechanical-electrical physical constraint model based on the current "fine screening condition" parameters. The model calculates the minute changes in the amplitude of the vibration table in real time based on the fluctuations in the current motor drive current, and simultaneously adjusts the shadow rendering of the virtual lighting to simulate the impact of lighting on the imaging of the vibrating silicon wafer in the real environment. At the same time, considering that the silicon wafer is extremely thin (e.g., 150μm thick) and has a large area, the system initiates a fluid-solid coupling mechanism to simulate the gas-solid coupling effect between the high-pressure airflow ejected from the screening nozzle and the high-speed moving silicon wafer. The twin accurately calculates the levitation force and lateral thrust of the airflow acting on the bottom of the silicon wafer, reconstructs the minute suspension height and tilt angle of the silicon wafer in the airflow field, and thus perfectly reproduces the entire physical process of the silicon wafer being lifted and suspended by the airflow at the fine screening station and imaged under specific lighting in the digital space.

[0024] Therefore, the digital twin undergoes multi-level iterations, successively going through primary iteration, intermediate iteration, and advanced iteration. At this stage, the primary iteration performs spatiotemporal registration and noise filtering to restore the real scene; the intermediate iteration uses implicit neural network technology to fill in image blind spots and deduce the invisible stress distribution inside the equipment; the advanced iteration introduces a long short-term memory mechanism to predict the future microsecond-level evolution of the equipment state based on the trajectory of previous operating conditions, thereby determining the current state of the screening equipment. The current state of the screening equipment not only includes the current working state of the screening equipment, but also embeds a tensor expression of the execution deviation, which is compatible with the overall consideration of predicting the future microsecond-level trajectory of previous operating conditions, ensuring the accuracy of the current state of the screening equipment.

[0025] At this stage, in the initial iteration phase, the digital twin primarily performs data cleaning and environmental restoration tasks. For the spatiotemporally aligned multi-source data, the system employs Kalman filtering or wavelet transform to smooth random noise in time-series data such as vibration frequency and transmission belt speed, eliminating high-frequency jitter interference caused by motor start-stop. Simultaneously, for image stream data, the system uses geometric registration to correct pixel-level distortion caused by camera angle jitter, dynamically calibrating the coordinate reference in virtual space with the real-time coordinates in the physical world at the microsecond level. This process aims to eliminate environmental interference factors, ensuring that the scene presented by the digital twin is a denoised "real scene," providing a high signal-to-noise ratio data foundation for subsequent in-depth analysis.

[0026] The intermediate iteration focuses on solving the perception problems of visual blind spots and internal invisible states. The system uses neural networks to learn the continuous feature representation of silicon wafers in multi-angle image streams, and performs pixel-level filling of silicon wafer edges or back areas that the visual array could not capture due to occlusion or depth of field limitations, reconstructing the complete three-dimensional global morphology of the silicon wafer. At the same time, the system combines an opto-mechanical-electrical physical constraint model and uses the concept of finite element analysis to map external vibration excitation and load conditions to the internal structure of the equipment, and deduce the invisible stress distribution and micro-deformation of key components such as bearings and cantilever beams, thereby achieving digital vision of mechanical fatigue and latent faults without disassembling the equipment.

[0027] In the advanced iteration stage, the system introduces a Long Short-Term Memory (LSTM) network architecture, utilizing its memory units to capture the long-range dependencies of the screening equipment in the time dimension. The system takes the operating conditions trajectory of the previous time step as the input sequence, such as the gradual change of vibration amplitude and the rise of motor temperature. Through a trained time-series prediction model, the system outputs the evolution trend of the equipment state within a future time window at the microsecond to millisecond level. Based on this prediction result, the system determines the "current state" of the screening equipment. This state is not a simple scalar set, but is encapsulated as a high-dimensional state vector, which specifically embeds a "tensor expression of execution deviation". This tensor mathematically represents the multi-dimensional error information such as the action lag, position offset, and force deviation of the actual actuator, intuitively reflecting the three-dimensional distribution of the control error of the equipment under the current operating condition.

[0028] Specifically, in the initial iteration, for silicon wafers running at high speed on the conveyor belt, the system detected high-frequency jagged artifacts at the image edges caused by vibration. The digital twin filtered out the vibration noise mixed in the image through wavelet threshold denoising and corrected the micro-displacement at the moment of camera exposure through geometric registration, so that the edges of the silicon wafer in virtual space were restored to a smooth and realistic geometric contour, restoring the real posture of the silicon wafer at the sorting station.

[0029] Following an intermediate iteration, the system detected that the lower right corner of the silicon wafer was in the camera's visual blind spot. Using implicit neural network technology, the system automatically generated high-fidelity pixel filling for the lower right corner blind spot based on the texture features and lighting patterns of other areas of the silicon wafer, eliminating the risk of missing a hidden crack in that area. At the same time, the system deduced the stress distribution inside the screening robot and found that under the long-term high-frequency grasping action of "fine screening", there was an implicit shear stress concentration of about 15MPa at the internal bearing connection of the robot joint. Although it did not reach the alarm threshold, it had been captured and recorded by the digital twin.

[0030] In the advanced iteration, the LSTM network analyzed the trajectory of the previous 10 seconds and predicted that within the next 500 microseconds, the end effector would experience a mechanical rebound hysteresis of approximately 0.02 mm due to the robot arm having just completed a heavy-load gripping operation. Based on this, the system determined the current state of the device and generated an execution deviation tensor. This tensor explicitly indicated a displacement deviation of -0.05 mm on the X-axis, an attitude deviation of 0.1° on the θ-axis, and a fluctuation deviation of 2 N in the end effector force. The state information of this embedded deviation tensor was then transmitted to the control system, providing a quantitative basis for subsequent precise correction.

[0031] In step S12, the specific steps are as follows: S121: Directional analysis is performed on the digital twin, and the image content of the silicon wafer is separated during the analysis process. The image content of the silicon wafer is dynamically identified at the pixel level to extract the micro-defects on the surface of the silicon wafer. The micro-defects on the surface of the silicon wafer are dynamically diffracted under illumination and motion blur, and image frames of the silicon wafer in different shapes are output. Multiple image frames are fused by multiple factors to generate a global feature map containing spatial topology and temporal evolution. Micro-defects include hidden cracks, micro-chipping, and textured color difference. S122: Obtain the convolutional neural network preset by the screening equipment, input the global feature map into the convolutional neural network, and activate different network branches according to the complexity of the global feature map: when the global feature map shows that the silicon wafer features are in a high-dimensional chaotic state, the convolutional neural network matches the fine screening condition; when the global feature map is clear and conforms to the standard template, the convolutional neural network reduces the dimension and matches the general screening condition; and for the global feature map with broken or severely distorted features, the convolutional neural network directly matches the waste rejection condition.

[0032] In the embodiments of this application, the digital twin is subjected to directional analysis, and the image content of the silicon wafer is separated during the analysis process. The image content of the silicon wafer is dynamically identified at the pixel level to extract the micro-defects on the surface of the silicon wafer. The micro-defects on the surface of the silicon wafer undergo dynamic diffraction under illumination and motion blur, and image frames of the silicon wafer in different shapes are output. Multiple image frames are fused by multiple factors to generate a global feature map containing spatial topology and temporal evolution. The micro-defects include hidden cracks, micro-chipping edges, and textured color differences.

[0033] At this point, the system performs deep-oriented analysis on the constructed digital twin to extract image content related to the silicon wafer. The system employs a deep learning-based semantic segmentation network to perform pixel-by-pixel fine deconstruction of the silicon wafer image and implements a dynamic recognition method. This method goes beyond static feature extraction and locks onto the microscopic areas of the silicon wafer surface, performing high-sensitivity tracking of hidden cracks, micro-chipping, and textured color differences. During the dynamic recognition process, the system fully considers the high-speed motion characteristics of the silicon wafer, focusing on the changes in image features between consecutive frames to ensure accurate locking of the pixel coordinates of defects in complex dynamic environments. Optionally, hidden cracks typically manifest as extremely fine linear changes in light intensity; micro-chipping presents as minute defects in edge geometry; and textured color differences present as uneven grayscale of surface texture.

[0034] To address the unique optical behavior of micro-defects under high-speed motion and specific lighting conditions, the system introduces a dynamic diffraction analysis mechanism. Since micro-defects on the silicon wafer surface undergo light scattering and diffraction under strong light, and motion blur alters the morphology of the diffraction fringes, an optical response model for the defects is established. This model simulates the dynamic diffraction effect produced by light at the defects, capturing the tailing and diffusion phenomena of defect features caused by motion blur. Based on this, the system outputs multi-morphological image frames of the silicon wafer under different lighting angles and at different motion times. These image frames not only contain the spatial geometric information of the defects but also implicitly contain the depth and structural information of the defects.

[0035] After acquiring multi-morphological image frames, the system performs a multi-factor fusion operation. Using a multi-scale feature fusion method, the defect features under different exposure conditions, different diffraction patterns, and different time slices are weighted and fused. The system constructs a global feature map containing spatial topology and temporal evolution. This map aggregates scattered pixel-level features into structured feature tensors, which fully characterize the surface state of the silicon wafer from entering the screening area to leaving it, providing high-dimensional feature input for subsequent convolutional neural network condition determination.

[0036] Specifically, the system analyzes a silicon wafer candidate that is being transmitted at a speed of 1.5 meters per second. Through ultra-high resolution pixel-level scanning, the system identifies a hidden crack with a length of only 50 micrometers in the lower left corner of the silicon wafer. At the same time, it detects a micro-chipping edge with a width of about 30 micrometers on the right edge and finds a tiny textured color difference in the center area caused by fluctuations in the texturing process. These micro-defects are extremely difficult to detect in a static state, but they are accurately marked in a dynamic identification method.

[0037] The system used an illumination model to analyze and found that when the silicon wafer moved directly under a line light source, the 50-micrometer hidden crack underwent significant dynamic diffraction under strong perpendicular light incidence. The light was scattered at the crack interface, causing the gray values ​​of the crack pixels in the image to exhibit a diffraction halo characteristic of "bright in the middle and dark on both sides". Under the influence of motion blur, the edge contour of the micro-cracked edge exhibited a specific trailing shape in the image frame. Based on this, the system output multiple sets of image data, including normal reflection frames, diffraction enhancement frames, and motion blur frames.

[0038] The system performs multi-factor fusion on these image frames of different forms, automatically extracts the optical scattering features of the crack in the diffraction frame, the edge trailing features of the broken edge in the blurred frame, and the color difference texture features in the normal frame, and maps them into a unified high-dimensional feature space. The final global feature map not only marks the specific spatial coordinates of the hidden crack, but also records the diffraction intensity change curve of the crack during the illumination change process.

[0039] Furthermore, the convolutional neural network preset by the screening equipment is obtained, and the global feature map is input into the convolutional neural network. The convolutional neural network activates different network branches according to the complexity of the global feature map: when the global feature map shows that the silicon wafer features are in a high-dimensional chaotic state, the convolutional neural network matches the fine screening condition; when the global feature map is clear and conforms to the standard template, the convolutional neural network reduces the dimension and matches the general screening condition; and for the global feature map with broken or severely distorted features, the convolutional neural network directly matches the waste rejection condition. The introduction of the convolutional neural network to directly match the waste rejection condition, along with the introduction of a digital twin, further controls the global feature map of the silicon wafer, improving the accuracy of the screening equipment in the silicon wafer screening condition.

[0040] At this point, the system retrieves a pre-trained convolutional neural network for silicon wafer sorting from the model library of the edge computing node. This network architecture is designed with a dynamic activation mechanism, which can adjust the computation path according to the characteristics of the input data. The system uses the global feature map generated in step S121 as the input tensor and performs feature extraction and encoding through convolutional layers. During this process, the network performs a depth scan of the feature map to analyze its feature space distribution density, texture entropy value, and edge gradient complexity, providing a quantitative basis for subsequent network branch activation.

[0041] Convolutional neural networks perform logical judgments and activate corresponding network branches based on the complexity index of the global feature map, specifically dividing them into three types of working condition matching paths: High-dimensional chaotic state matching fine screening conditions: When the network detects that the entropy value of the feature map is high, presenting a "high-dimensional chaotic state", it means that there are a lot of fine textures or suspected defect features on the surface of the silicon wafer, such as diffraction spots generated by hidden cracks. Although these features do not constitute a clear waste signal, they require extremely high precision identification. At this time, the network automatically activates the deep residual branch, uses high convolution kernel density to perform refined feature extraction, and outputs fine screening condition instructions, requiring the equipment to perform low-speed, high-precision sorting. Clear standard matching screening conditions: When the feature map outline is clear, the gray level distribution is uniform, and the cross-correlation coefficient with the standard good template is higher than the set threshold, it indicates that the silicon wafer quality is stable; the network activates the lightweight branch to perform rapid dimensionality reduction processing, directly mapping to the screening conditions, allowing the equipment to operate at standard efficiency; Fracture distortion matching waste rejection condition: When the feature map shows obvious geometric fracture, severe edge distortion or large area of ​​texture loss, the network judges it as structural damage, skips the conventional classification logic, and directly maps it to the waste rejection condition.

[0042] Specifically, the system inputs the previously generated global feature map containing hidden crack diffraction features and micro-edge chipping and tailing features into a preset convolutional neural network. The network encodes the features of the map and finds that the texture entropy value of the map is significantly higher than that of conventional silicon wafers. Moreover, the feature space is filled with fine scattered light spots and non-uniform textures, exhibiting a typical "high-dimensional chaotic state". This is because the criteria for judging silicon wafers are extremely strict. Tiny hidden cracks and color differences are manifested as complex non-linear distributions in the feature space, rather than simple clear outlines.

[0043] Based on this complexity analysis, the convolutional neural network determines that although the candidate silicon wafer has complex features, it is not structurally damaged and requires further refinement. The network then activates a deep residual network branch to perform layer-by-layer convolution and pooling operations on the graph, extracting high-dimensional feature vectors and finally outputting the classification result of the fine screening condition. This decision means that the screening equipment needs to start the "fine screening mode" for the silicon wafer to distinguish it from the ordinary screening process for qualified products, thereby ensuring that the silicon wafer is not misjudged as a common product and guaranteeing the output rate of high-value products. Conversely, if the input graph shows that there is a large area of ​​fracture in the center of the silicon wafer, the network will directly classify it as a waste rejection condition.

[0044] In step S13, the specific steps are as follows: S131: The sieving conditions of the silicon wafers by the sieving equipment and the current state of the sieving equipment are input to a unified control space through different channels. In this control space, the sieving conditions of the silicon wafers by the sieving equipment and the current state of the sieving equipment are projected to the same topological manifold through manifold learning, and the fuzzy PID control mechanism is triggered synchronously in the conflict detection. S132: In the fuzzy PID control mechanism, fuzzy logic is used to transform the risk information of the silicon wafer into the offset trend of the PID parameters. At the same time, the integral part of the PID introduces the fatigue decay factor of the current state of the equipment, while the derivative part is coupled with the gradient change rate of the global feature map of the silicon wafer, thereby determining the risk information of the silicon wafer. S133: When risk information of silicon wafers is detected, the corresponding response signal is acquired, and the screening equipment is triggered to adaptively adjust the screening conditions of silicon wafers along the response signal. The nonlinear collaborative adjustment information of the screening equipment in different working dimensions is output to realize the adaptive and accurate mapping between silicon wafers of different product grades and the operating parameters of the screening equipment.

[0045] In the embodiments of this application, the sieving conditions of the silicon wafers by the sieving equipment and the current state of the sieving equipment are input to a unified control space through different channels. In this control space, the sieving conditions of the silicon wafers by the sieving equipment and the current state of the sieving equipment are projected onto the same topological manifold through manifold learning, and a fuzzy PID control mechanism is synchronously triggered in the conflict detection, thus introducing a fuzzy PID control mechanism that is synchronously triggered in the conflict detection.

[0046] At this point, the system constructs a unified control space with multiple input interfaces, which serves as the hub connecting the perception and decision-making layer and the execution and control layer. The screening conditions of the silicon wafers and the current state of the screening equipment are defined as two independent input sources, which are transmitted in parallel to the control space along isolated data channels. In this process, the screening conditions, as "task instructions," represent the expected screening objectives, such as the low-speed and stable environment required for fine screening, while the current state of the screening equipment, as "constraints," represents the physical capability boundaries of the equipment, such as actual vibration deviations or mechanical hysteresis.

[0047] Since screening conditions and equipment status are heterogeneous data, direct fusion would lead to the curse of dimensionality. The system introduces a manifold learning approach to construct a high-dimensional feature mapping model in the control space. Through local linear embedding or isometric mapping, the system projects the high-dimensional "screening conditions" and "equipment status" onto the same low-dimensional topological manifold. In this manifold space, the originally discrete condition points and status points are transformed into continuous geometric trajectories. The system calculates the Euclidean distance or geodesic distance between the two in the manifold space to quantitatively evaluate the degree of matching between the "target condition" and the "current status," thereby providing a geometrically intuitive decision basis for the generation of control strategies.

[0048] Within the topological manifold space, the system implements real-time conflict detection logic. If the manifold projection result shows a large spatial deviation between the "task target point" and the "equipment status point," meaning that the current state of the equipment cannot meet the operating environment requirements (e.g., the fine screening requires extremely low vibration, but the equipment status display shows high-frequency vibration), it is determined to be a "state-condition conflict." Once such a conflict or deviation exceeds the threshold, the system immediately triggers the fuzzy PID control mechanism. This mechanism does not directly execute rigid control but uses fuzzy logic to fuzzify the distance deviation and deviation change rate in the manifold space, transforming them into fuzzy sets such as "large deviation" and "small deviation." Based on the fuzzy rule base, the dynamic correction parameters of the PID controller are calculated in real time, thereby achieving smooth and adaptive adjustment of the operating state of the screening equipment.

[0049] Specifically, the system inputs the "fine screening condition" command output by the decision layer and the current equipment status detected in S113 into the unified control space respectively; the "fine screening condition" command requires the conveyor belt speed to be reduced to 1.0 m / s and the vibration frequency to be adjusted to 30 Hz to achieve stable conveying; the current equipment status includes the execution deviation tensor, and the vibration table has micron-level low-frequency shaking.

[0050] The system maps these two heterogeneous data to the same topological manifold. In this manifold space, the "fine screening condition" is mapped to an ideal stable pole, while the "current state of the equipment" is mapped to a dynamic trajectory point that deviates from the pole. There is a significant spatial distance between the two points, which indicates that the current physical vibration and rotational speed deviation of the equipment cannot meet the ultra-stable environment required for silicon wafer fine screening.

[0051] The system detects this conflict in the manifold space: the device state point has not yet converged to the neighborhood of the target point. This conflict signal immediately triggers the fuzzy PID control mechanism. Based on the distance vector in the manifold space, the fuzzy controller analyzes that the speed deviation is "positive and large," the actual speed is too fast, and the deviation change rate is "negative and small," indicating that it is decelerating but not enough. According to the fuzzy rule base, such as the rule: IF deviation is positive and large AND change rate is negative and small THEN increase the proportional coefficient, the system dynamically adjusts the PID control parameters, outputs correction commands, and drives the motor to quickly and smoothly adjust the speed and vibration to the precise range required for the fine screening condition, eliminating the deviation between the state and the target, and creating physical conditions for the precise grasping of silicon wafers.

[0052] Furthermore, in the fuzzy PID control mechanism, fuzzy logic is used to transform the risk information of the silicon wafer into the offset trend of the PID parameters. At the same time, the integral term of the PID introduces the fatigue decay factor of the current state of the equipment, while the derivative term couples the gradient change rate of the global feature map of the silicon wafer, thereby determining the risk information of the silicon wafer. This approach takes into account the overall consideration of the gradient change rate of the global feature map of the silicon wafer, ensuring the accuracy of the risk information of the silicon wafer.

[0053] At this point, at the core layer of the fuzzy PID control mechanism, the system constructs a fuzzy inference engine to fuzzify the risk information of the silicon wafer. The system defines fuzzy linguistic variables to map specific numerical risk indicators to fuzzy sets, such as "extremely high risk", "medium risk", and "extremely low risk". Based on the preset fuzzy control rule table, the system infers the parameter adjustment trend of the PID controller. Specifically, when the risk information indicates that the silicon wafer is in a high-risk state, the fuzzy logic outputs a "negative offset" instruction, tending to increase the proportional coefficient to enhance the rigidity of the system response, or decrease the derivative coefficient to avoid mechanical shock caused by over-adjustment, thereby adapting to changes in risk level through dynamic parameter offset.

[0054] To address the issues of integral saturation and control rigidity in traditional PID integral terms, a fatigue attenuation factor based on the current state of the equipment is introduced into the integral term. This factor is a real-time scalar based on the output of step S113 of the digital twin, reflecting the fatigue level and performance degradation of key components of the equipment. The system uses this factor as a weighting coefficient for the integral gain to construct an adaptive integral term. When the equipment is in a high fatigue state, the attenuation factor reduces the intensity of the integral action to prevent integral overshoot and system oscillation caused by equipment response lag. Conversely, when the equipment is in good condition, the standard integral intensity is maintained to eliminate steady-state error. This mechanism ensures that the control strategy can adapt to the health state of the equipment, achieving precise control that "acts within its capabilities." Key components of the equipment include bearings and conveyor belts.

[0055] To enhance the system's instantaneous response to changes in silicon wafer morphology, the differential link is designed to be deeply coupled with the aforementioned generated global feature map. The system calculates the gradient change rate of key features in the global feature map in real time. This gradient change rate characterizes the abrupt change trend of the silicon wafer's appearance during dynamic transmission. This gradient change rate is superimposed on the differential link as a feedforward signal, enabling the controller not only to sense the rate of change in the device's motion state but also to predict abrupt changes in the silicon wafer's appearance, such as a sudden transition from a intact area to a chipped area. Through this cross-domain coupling of opto-mechanical-electrical components, the differential link can output a damping signal in advance to suppress control lag or mechanical jitter caused by abrupt changes in the silicon wafer's morphology. Key features include edge contours and defect textures.

[0056] Specifically, the system detects the presence of "hidden cracks" in the global feature map of a silicon wafer. The fuzzy inference engine interprets this high-risk feature as a "large negative offset" trend in the PID parameters, which manifests as an instantaneous increase in the proportional coefficient. This improves the response speed of the screening robot or rejection mechanism, ensuring that the silicon wafer is quickly adjusted before entering the fine screening and grasping area, thus avoiding the hidden cracks from breaking due to slow action.

[0057] In the integral processing, considering that the screening equipment has been running continuously for more than 200 hours, the digital twin detected slight wear on the bearing of the vibrating motor. Based on this, the system generates a fatigue attenuation factor less than 1 and multiplies it into the integral term calculation formula. This results in a moderate reduction in the integral effect, thereby effectively avoiding the integral oversaturation phenomenon caused by the hysteresis effect due to the mechanical wear of the equipment under the control requirements of low speed and high precision screening, and ensuring the stability of the conveyor belt speed control.

[0058] In the differential coupling, when the silicon wafer passes through the visual inspection area, the micro-edge chipping features of its edge manifest as a sharp gradient jump in the global feature map. The differential link captures this gradient change rate in real time and immediately outputs a predictive damping adjustment signal. This allows the actuator of the sorting equipment to fine-tune the contact force of the pneumatic nozzle or the posture of the robotic arm before it physically contacts the silicon wafer, thereby achieving a closed-loop response to risk information while ensuring the integrity of the silicon wafer.

[0059] Therefore, when risk information of silicon wafers is detected, the corresponding response signal is obtained, and the screening equipment is triggered to adaptively adjust the screening conditions of silicon wafers along the response signal. The nonlinear collaborative adjustment information of the screening equipment in different working dimensions is output, so as to realize the adaptive and precise mapping between silicon wafers of different product grades and the operating parameters of the screening equipment.

[0060] At this time, the system monitors the control variable output by the fuzzy PID control mechanism in real time. When the variable exceeds the preset safety threshold or a specific logic range, it is determined that risk information of the silicon wafer has been detected, such as the risk of hidden crack propagation or the risk of micro-edge breakage. The system then captures the state variable as the corresponding response signal, which is essentially a trigger pulse containing the risk level and characteristic type. Along the transmission path of the response signal, the system forcibly interrupts or corrects the current standard screening logic, triggers the adaptive adjustment mechanism of the screening equipment for the silicon wafer screening conditions, and switches the control mode from "normal following" to "risk active adaptation" state.

[0061] For the complex electromechanical coupling system of screening equipment, linear adjustment of a single parameter is often insufficient to mitigate risks. The system constructs a multi-dimensional control matrix, outputting nonlinear collaborative adjustment information in different working dimensions. Specifically, the system calculates in parallel the adjustment amounts of vibration frequency, conveyor belt speed, sorting actuator force, and airflow intensity. These adjustment information do not exist in isolation but are interconnected through nonlinear coupling: for example, while reducing the conveying speed, the vibration phase is nonlinearly changed to counteract inertial jitter; while increasing the sorting gripping force, the pneumatic suspension pressure is adjusted inversely to maintain the stability of the silicon wafer posture. Through this multi-dimensional, nonlinear collaborative adjustment, the dynamic stability of the equipment in dealing with risks is ensured.

[0062] After completing the coordinated adjustment, the system establishes a dynamic mapping function in the control space based on the current silicon wafer risk characteristics and equipment status. This function breaks the fixed parameter configuration table and realizes the real-time correspondence between silicon wafers of different product grades and the operating parameters of the screening equipment. The system writes the adjusted set of operating parameters into the PLC controller, so that the equipment can have a set of exclusive optimal operating parameters for each silicon wafer with specific risk characteristics, such as different product grades such as Grade A, Grade B, and breakage grade; the operating parameters of the screening equipment include the final rotation speed, frequency, and force after PID correction.

[0063] Specifically, in the risk response and triggering phase, when the system detects that a certain silicon wafer has a "hidden crack" risk, the fuzzy PID controller outputs a response signal that indicates "fragile and high risk". This signal instantly triggers the adaptive adjustment logic of the screening equipment. The system determines that the current conventional fine screening parameters may cause the hidden crack to be stressed and expanded, and must be corrected immediately.

[0064] The system outputs multi-dimensional adjustment commands; specifically, in terms of speed, the system non-linearly reduces the conveyor belt speed to 0.8 m / s, and non-linearly decelerates to avoid inertial impact caused by sudden stops; in terms of mechanics, it synchronously adjusts the pneumatic end effector of the sorting robot arm, and in conjunction with the sorting airflow in the fluid dimension, it lowers the bottom suspended air pressure to match the mass change after the silicon wafer cracks, preventing the airflow from being too large and blowing the silicon wafer apart. These adjustment information are sent out synchronously within milliseconds, forming a multi-physics field collaborative protection.

[0065] The system defines this silicon wafer with "hidden cracks" as a "Grade A - Vulnerable" product and binds the adjusted operating parameters to it. The sorting equipment performs its actions according to this mapping relationship, successfully sorting this silicon wafer, which has minor flaws but is still of high value, into the premium material box without damage, thus avoiding silicon wafer breakage caused by rigid equipment parameters.

[0066] In step S14, the specific steps are as follows: S141: Obtain the adjusted screening conditions, perform reverse tracing based on these conditions, and combine the feature extraction mechanism during the tracing process to extract multiple online operating parameters of the screening equipment; a multi-source heterogeneous vision array performs re-inspection at the screening discharge port, outputs high-confidence screening results in real time, and compares the screening results in time, space, and causal dimensions using multi-dimensional tensors to determine screening deviation information with deep physical meaning; S142: Input the screening deviation information into the blockchain space, and gradually build a decentralized multi-condition screening system by combining the corresponding blockchain technology: In this multi-condition screening system, the screening deviation information, screening conditions and online working parameters are packaged into blocks, the authenticity of the deviation is verified by the on-chain consensus mechanism, and the parameter correction instructions of the digital twin and convolutional neural network are generated through the on-chain preset reverse calculation contract.

[0067] In the embodiments of this application, the adjusted screening conditions are obtained, and reverse tracing is performed based on these conditions. During the tracing process, a feature extraction mechanism is combined to extract multiple online operating parameters of the screening equipment. A multi-source heterogeneous vision array performs a re-inspection at the screening discharge port and outputs high-confidence screening results in real time. The screening results are placed in the time dimension, spatial dimension, and causal dimension for multi-dimensional tensor comparison to determine screening deviation information with deep physical meaning. This approach incorporates the overall consideration of the time dimension, spatial dimension, and causal dimension, ensuring the accuracy of screening deviation information with deep physical meaning.

[0068] At this point, the system locks the final screening condition label after adaptive adjustment, using it as the logical starting point for tracing. The system constructs a reverse propagation path, tracing back along the logical link of "condition label - control command - actuator". During this process, the system embeds a feature extraction mechanism to accurately extract the core online working parameters that determine the condition from the massive real-time data stream. These parameters include, but are not limited to, real-time vibration frequency amplitude, instantaneous acceleration of the end effector, pneumatic pressure value, and real-time linear speed of the conveyor belt. These parameters are encapsulated into parameter vectors, reflecting the actual physical input of the equipment when performing a specific screening action.

[0069] At the end of the screening process, the system deploys a multi-source heterogeneous vision array for secondary inspection. Unlike the dynamic imaging at the screening station, the vision array here focuses on result verification. It uses a high frame rate area array camera and a structured light sensor to perform static or low-speed dynamic re-inspection on the silicon wafers that have fallen into the material box or conveyor belt. The system eliminates ambient light interference and occlusion noise through multi-view geometric reconstruction and image fusion to generate high-confidence screening results. These results include not only the final appearance of the silicon wafers but also their final physical position and orientation in the material box, objectively verifying the effectiveness of the screening action.

[0070] The system constructs two high-dimensional tensors: "expected screening target: prediction from the digital twin" and "actual screening result: re-inspection from the vision array," and performs multi-dimensional comparison operations. In the time dimension, it compares the difference between the execution sequence of the screening actions and the expected sequence to identify whether the actions are lagging or ahead. In the spatial dimension, it calculates the Euclidean distance deviation between the actual wafer placement position and the target position. In the causal dimension, it analyzes the physical correlation between the input parameters and the output results. Through this multi-dimensional tensor comparison, the system not only outputs numerical errors but also analyzes the deep physical causes of these errors, such as "slippage caused by micro-vibration" or "attitude rollover caused by residual wind," thereby determining screening deviation information with deep physical significance.

[0071] Specifically, the system acquired the current "A-level fine screening flexible gripping" working condition; the system traced back to the control layer and extracted that the actual gripping speed of the screening robot arm at this time was 0.5 m / s, which was 0.2 m / s lower than the normal speed, the negative pressure value of the end suction nozzle was -45 kPa, which was in the low-pressure flexible range, and the frequency of the vibration transmission table was maintained at 30 Hz. These parameters were extracted as benchmark data to verify whether the screening action was compliant.

[0072] After the silicon wafer falls into the premium material box after being sieved, the multi-source heterogeneous vision array located above the material inlet immediately captures an image of it. The system then uses image processing to confirm that the silicon wafer has successfully fallen into the designated slot and that there are no new mechanical scratches or chipping on the surface. The system outputs a "sieving success" result with a confidence level of 99.8%.

[0073] The system performs a deep comparison between the re-inspection results and the expected target. In terms of time, the system found that the robotic arm's material release time was 10 milliseconds later than expected. In terms of space, the actual position of the silicon wafer in the material box was measured to be 1.5 millimeters off from the center point. In terms of causality, the system correlated and analyzed the previously extracted parameter "vibration frequency 30Hz" and found that the offset was caused by a small slippage due to the accumulation of micro-resonance in the material box caused by the 30Hz vibration frequency over a long period of time. Based on this, the system determined the deeper physical meaning of the screening deviation: "low-frequency vibration coupling under fine screening conditions caused the silicon wafer drop position to drift," and generated feedback instructions accordingly, providing a precise basis for subsequent adjustment of the material box's vibration reduction parameters.

[0074] Furthermore, the screening deviation information is input into the blockchain space, and a decentralized multi-condition screening system is gradually constructed by combining the corresponding blockchain technology. In this multi-condition screening system, the screening deviation information, screening conditions, and online working parameters are packaged into blocks. The authenticity of the deviation is verified by the on-chain consensus mechanism, and the parameter correction instructions of digital twins and convolutional neural networks are generated through the on-chain preset reverse calculation contract. The parameter correction instructions of digital twins and convolutional neural networks are introduced. At the same time, the adaptive mapping between silicon wafers of different product grades and the operating parameters of screening equipment is realized. By fully considering the screening deviation information and the blockchain, the accuracy of the multi-condition screening system is improved.

[0075] At this point, the system constructs a blockchain underlying architecture for the Industrial Internet of Things (IIoT). The screening deviation information, corresponding screening condition tags, and real-time online working parameters obtained in step S141 are serialized according to a preset data structure. The system packages these multi-source heterogeneous data into a standardized data block. This block contains the key hash value, timestamp, and index information of the previous block for the entire screening process of the current batch of silicon wafers. The data is permanently recorded on the distributed ledger through encryption, ensuring the immutability and traceability of the screening process data and providing a data foundation for the establishment of subsequent trust mechanisms.

[0076] During block broadcasting, the system triggers the on-chain consensus mechanism, where multiple verification nodes in the blockchain network cross-verify the broadcast screening deviation information. The nodes compare the deviation information with the historical operating data and equipment status logic recorded on the chain to verify whether the deviation is caused by a real physical fault or environmental interference, rather than a sensor false alarm or malicious attack. Only when a majority of nodes reach a consensus and confirm the authenticity and validity of the deviation data is the block officially added to the end of the blockchain, completing the transformation from "data record" to "credible fact". The multiple verification nodes include edge computing nodes and quality monitoring nodes.

[0077] After successful verification, the system automatically triggers a reverse computation smart contract deployed on the blockchain. This contract incorporates a parameter optimization method, which calculates the optimal parameter configuration required to eliminate the confirmed screening deviation based on the confirmed screening deviation information. The contract outputs parameter correction instructions for the digital twin and convolutional neural network (CNN). Specifically, the contract calculates the model weight update amount that can compensate for physical deviations, generates physical parameter calibration instructions for the digital twin, and hyperparameter adjustment instructions for the CNN model, thereby achieving self-evolution and closed-loop optimization of the screening system. Physical parameter calibration instructions include correcting the fluid coupling coefficient; hyperparameter adjustment instructions for the CNN model include adjusting the confidence threshold for defect identification.

[0078] Specifically, the system encapsulates the screening deviation information detected in S141, namely "silicon wafer drop position drift of 1.5 mm", along with the online working parameters such as "fine screening flexible grasping condition", actual speed of the robotic arm: 0.5 m / s, and vibration frequency, into a new data block. This block is then encrypted by hash and broadcast to the blockchain network of the entire screening workshop as the unique digital information for the entire silicon wafer screening process.

[0079] The quality verification nodes and device status monitoring nodes in the blockchain network receive the block; the nodes execute the consensus verification logic: check whether the 1.5 mm displacement deviation at a vibration frequency of 30 Hz conforms to the physical law of mechanical resonance; through cross-comparison of historical data by multiple nodes, it is confirmed that the deviation is actually caused by the micro-vibration of the device rather than the random error of the visual detection system, thus reaching a consensus, confirming the validity of the deviation information, and permanently storing it on the chain.

[0080] During the reverse calculation and correction phase, the on-chain pre-set reverse calculation smart contract is automatically triggered. The contract reverse-engineers the "material drop drift" problem, calculating that to eliminate the deviation, the vibration damping coefficient in the digital twin needs to be increased by 0.05. It also suggests that the convolutional neural network should increase the sensitivity weight of vibration spectrum features when judging the fine screening conditions. Based on this, the contract generates correction instructions: on the one hand, it instructs the digital twin to update its vibration model parameters to improve simulation accuracy; on the other hand, it fine-tunes the network weights of the CNN so that when it encounters similar vibration spectra in the future, it can more accurately predict the material drop deviation and compensate for it in advance at the control level, thereby constructing a decentralized multi-condition screening system with self-learning and self-correction capabilities.

[0081] Please see Figure 2 , Figure 2 This is a schematic diagram of the structural composition of a multi-condition screening system for screening equipment according to an embodiment of the present invention; the multi-condition screening system is applied to the above-mentioned multi-condition screening method for screening equipment; the multi-condition screening system for screening equipment includes: The digital twin module 21 is used to collect multiple working data of the screening equipment at the current moment, and combine them with multiple image data of the screening equipment at the screening station to create a digital twin to construct a corresponding digital twin; the current state of the screening equipment is determined based on the multi-level iteration of the digital twin. The working condition module 22 is used to determine the image content of each silicon wafer based on the analysis of the digital twin, determine the global feature map of the silicon wafer by dynamically recognizing the image content, and further combine the pre-set convolutional neural network of the screening equipment to determine the screening working condition of the screening equipment for the silicon wafer. The screening working condition covers fine screening working condition, general screening working condition, and waste removal working condition. The adaptive adjustment module 23 is used to input the screening conditions of the screening equipment for silicon wafers and the current state of the screening equipment into a unified control space, and trigger a fuzzy PID control mechanism in the control space. This triggers the adaptive adjustment of the screening conditions of the screening equipment for silicon wafers when risk information of silicon wafers is detected, thereby realizing the adaptive mapping between silicon wafers of different product grades and the operating parameters of the screening equipment. The multi-condition screening system module 24 is used to mark the adjusted screening conditions, determine multiple online working parameters of the screening equipment based on the traceability of the screening conditions, and output the corresponding screening results in real time in combination with the multi-source heterogeneous vision array. Based on the comparison of the screening results in different dimensions, the corresponding screening deviation information is determined, and the corresponding multi-condition screening system is constructed in combination with blockchain.

[0082] 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.

[0083] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein; this application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein; the specification and embodiments are to be considered exemplary only.

[0084] 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. A screening method based on multiple working conditions for a screening device, characterized in that, include: Collect multiple working data of the screening equipment at the current moment, and combine them with multiple image data of the screening equipment at the screening station to create a digital twin, so as to construct a corresponding digital twin; determine the current state of the screening equipment based on the multi-level iteration of the digital twin; Based on the analysis of the digital twin, the image content of each silicon wafer is determined. The global feature map of the silicon wafer is determined by the dynamic recognition of the image content. Furthermore, the screening conditions of the screening equipment for silicon wafers are determined by the pre-set convolutional neural network of the screening equipment. The screening conditions include fine screening, general screening, and waste removal. The screening conditions of the silicon wafers and the current state of the screening equipment are input into a unified control space, and a fuzzy PID control mechanism is triggered in the control space. When risk information of silicon wafers is detected, the screening conditions of the screening equipment are adaptively adjusted, thereby realizing the adaptive mapping between silicon wafers of different product grades and the operating parameters of the screening equipment. The adjusted screening conditions are marked, and multiple online operating parameters of the screening equipment are determined based on the traceability of the screening conditions. The corresponding screening results are output in real time by combining the multi-source heterogeneous vision array. The corresponding screening deviation information is determined by comparing the screening results in different dimensions, and the corresponding multi-condition screening system is constructed by combining blockchain.

2. The screening equipment according to claim 1, based on a multi-condition screening method, is characterized in that, The collection and sieving equipment collects multiple working data at the current moment and combines them with multiple image data of the sieving equipment at the sieving station to create a digital twin, thereby constructing a corresponding digital twin. The current state of the screening equipment is determined based on the multi-level iterations of the digital twin, including: The screening equipment is monitored in real time, and multiple working data of the screening equipment at the current moment are determined based on the detection of the screening equipment database. These multiple working data cover vibration frequency, conveyor belt speed, execution actions, and dynamic detection data. The multi-source heterogeneous vision array of the screening equipment is marked, and the silicon wafers in the screening area are dynamically photographed based on the multi-source heterogeneous vision array. Multiple image data of the silicon wafers under high-speed movement are presented as an image stream of the silicon wafers at different angles.

3. The screening equipment according to claim 2, based on a multi-condition screening method, is characterized in that, The collection and sieving equipment collects multiple working data at the current moment and combines them with multiple image data of the sieving equipment at the sieving station to create a digital twin, thereby constructing a corresponding digital twin. Determining the current state of the screening equipment based on the multi-level iterations of the digital twin also includes: Spatiotemporal alignment of multiple working data and multiple image data is performed, and a digital twin mechanism is introduced in the alignment process to dynamically construct a digital twin with fluid-solid coupling and full-link physical constraints of optical-mechanical-electrical systems; The digital twin undergoes multiple iterations, successively going through primary iteration, intermediate iteration, and advanced iteration. At this stage, the primary iteration performs spatiotemporal registration and noise filtering to restore the real scene. The intermediate iteration uses implicit neural technology to fill in image blind spots and infer the invisible stress distribution inside the equipment. The advanced iteration introduces a long short-term memory mechanism to predict the future microsecond-level evolution of the equipment state based on the trajectory of previous operating conditions, thereby determining the current state of the screening equipment. The current state of the screening equipment not only includes the current working state of the screening equipment, but also embeds a tensor expression of the execution deviation.

4. The screening method based on multiple working conditions of the screening equipment according to claim 1, characterized in that, The process involves determining the image content of each silicon wafer based on the analysis of the digital twin, determining the global feature map of the silicon wafer by dynamically recognizing the image content, and further combining this with the convolutional neural network preset by the screening equipment to determine the screening conditions of the silicon wafers. These screening conditions include fine screening, general screening, and waste removal, among others. The digital twin is subjected to directional analysis, and the image content of the silicon wafer is separated during the analysis process. The image content of the silicon wafer is dynamically identified at the pixel level to extract the micro-defects on the silicon wafer surface. The micro-defects on the silicon wafer surface are dynamically diffracted under illumination and motion blur, and image frames of the silicon wafer in different shapes are output. Multiple image frames are fused by multiple factors to generate a global feature map containing spatial topology and temporal evolution. The micro-defects include hidden cracks, micro-chipping edges, and textured color difference.

5. The screening equipment according to claim 4, based on a multi-condition screening method, is characterized in that, The process involves determining the image content of each silicon wafer based on the analysis of the digital twin, determining the global feature map of the silicon wafer by dynamically recognizing the image content, and further combining this with the convolutional neural network preset by the screening equipment to determine the screening conditions of the silicon wafers. These screening conditions include fine screening, general screening, and waste removal, and also include: The convolutional neural network preset by the screening equipment is obtained, and the global feature map is input into the convolutional neural network. The convolutional neural network activates different network branches according to the complexity of the global feature map: when the global feature map shows that the silicon wafer features are in a high-dimensional chaotic state, the convolutional neural network matches the fine screening condition; when the global feature map is clear and conforms to the standard template, the convolutional neural network reduces the dimension and matches the general screening condition; and for the global feature map with broken or severely distorted features, the convolutional neural network directly matches the waste rejection condition.

6. The screening equipment according to claim 1, based on a multi-condition screening method, is characterized in that, The process of inputting the sieving conditions of the silicon wafers by the sieving equipment and the current state of the sieving equipment into a unified control space, and triggering a fuzzy PID control mechanism in this control space, thereby triggering adaptive adjustment of the sieving conditions of the silicon wafers by the sieving equipment when risk information of the silicon wafers is detected, includes: The sieving conditions of the silicon wafers by the sieving equipment and the current state of the sieving equipment are input to a unified control space through different channels. In this control space, the sieving conditions of the silicon wafers by the sieving equipment and the current state of the sieving equipment are projected to the same topological manifold through manifold learning, and the fuzzy PID control mechanism is triggered synchronously in the conflict detection. In the fuzzy PID control mechanism, fuzzy logic is used to transform the risk information of the silicon wafer into the offset trend of the PID parameters. At the same time, the integral term of the PID introduces the fatigue decay factor of the current state of the equipment, while the derivative term is coupled with the gradient change rate of the global feature map of the silicon wafer, thereby determining the risk information of the silicon wafer.

7. The screening method based on multiple working conditions of the screening equipment according to claim 6, characterized in that, The step of inputting the sieving conditions of the silicon wafers by the sieving equipment and the current state of the sieving equipment into a unified control space, and triggering a fuzzy PID control mechanism in this control space, so as to trigger the adaptive adjustment of the sieving conditions of the silicon wafers by the sieving equipment when risk information of the silicon wafers is detected, further includes: When risk information about silicon wafers is detected, the corresponding response signal is acquired. The sieving equipment is then triggered to adaptively adjust its sieving conditions along the response signal. The nonlinear collaborative adjustment information of the sieving equipment in different working dimensions is output, thereby achieving an adaptive and precise mapping between silicon wafers of different product grades and the operating parameters of the sieving equipment.

8. The screening equipment according to claim 1, based on a multi-condition screening method, is characterized in that, The adjusted screening conditions, based on the traceability of these conditions, determine multiple online operating parameters of the screening equipment. Combined with a multi-source heterogeneous vision array, the corresponding screening results are output in real time. Based on the comparison of these results across different dimensions, corresponding screening deviation information is determined. Furthermore, a multi-condition screening system is constructed using blockchain technology, including: The adjusted screening conditions are obtained, and reverse tracing is performed based on these conditions. During the tracing process, a feature extraction mechanism is combined to extract multiple online operating parameters of the screening equipment. A multi-source heterogeneous vision array performs a re-inspection at the screening discharge port and outputs high-confidence screening results in real time. The screening results are placed in the time dimension, spatial dimension, and causal dimension for multi-dimensional tensor comparison to determine screening deviation information with deep physical meaning.

9. The screening method based on multiple working conditions of the screening equipment according to claim 8, characterized in that, The adjusted screening conditions, based on the traceability of these screening conditions, determine multiple online operating parameters of the screening equipment, and combine this with a multi-source heterogeneous vision array to output the corresponding screening results in real time. The corresponding screening deviation information is determined based on the comparison of these screening results across different dimensions, and a corresponding multi-condition screening system is constructed using blockchain technology. The system also includes: The screening deviation information is input into the blockchain space, and a decentralized multi-condition screening system is gradually built by combining the corresponding blockchain technology. In this multi-condition screening system, the screening deviation information, screening conditions and online working parameters are packaged into blocks, the authenticity of the deviation is verified by the on-chain consensus mechanism, and the parameter correction instructions of the digital twin and convolutional neural network are generated through the on-chain preset reverse calculation contract.

10. A screening device based on a multi-condition screening system, characterized in that, The screening equipment based on the multi-condition screening system is applied to the screening equipment based on the multi-condition screening method as described in any one of claims 1-9; The screening equipment is a multi-condition screening system including: The digital twin module is used to collect multiple working data of the screening equipment at the current moment, and combine them with multiple image data of the screening equipment at the screening station to create a digital twin to construct a corresponding digital twin; the current state of the screening equipment is determined based on the multi-level iteration of the digital twin. The working condition module is used to determine the image content of each silicon wafer based on the analysis of the digital twin, determine the global feature map of the silicon wafer by dynamically recognizing the image content, and further combine the pre-set convolutional neural network of the screening equipment to determine the screening working condition of the screening equipment for the silicon wafer. The screening working condition covers fine screening working condition, general screening working condition, and waste removal working condition. The adaptive adjustment module is used to input the screening conditions of the silicon wafers and the current state of the screening equipment into a unified control space, and trigger a fuzzy PID control mechanism in the control space. This triggers the adaptive adjustment of the screening conditions of the silicon wafers when risk information of the silicon wafers is detected, thereby realizing the adaptive mapping between the operating parameters of silicon wafers of different product grades and the screening equipment. The multi-condition screening system module is used to mark the adjusted screening conditions, determine multiple online working parameters of the screening equipment based on the traceability of the screening conditions, and output the corresponding screening results in real time in combination with a multi-source heterogeneous vision array. Based on the comparison of the screening results in different dimensions, the corresponding screening deviation information is determined, and the corresponding multi-condition screening system is constructed in combination with blockchain.

Images