An intelligent control method for wafer cleaning and conveying process

Abstract

The present invention relates to the field of information technology, and in particular to an intelligent control method for a wafer cleaning and conveying process of semiconductor wafer cleaning technology. The method comprises: during the wafer conveying process, using a gyroscope sensor to detect vibration and displacement signals of a conveying device in real time, inputting the collected signals into a pre-trained support vector machine model for classification, and judging whether the current vibration and displacement exceed a preset threshold. The support vector machine model is trained using historical data and can accurately distinguish between normal and abnormal vibration and displacement states; when vibration or displacement exceeding the threshold is detected, triggering a servo motor for compensation control, suppressing vibration and displacement by adjusting conveying speed and acceleration parameters, and ensuring stable conveyance of the wafer to a rotating cleaning device. The compensation control algorithm dynamically adjusts control parameters according to the amplitude of vibration and displacement to achieve rapid response and precise compensation.

Description

Technical Field

[0001] The present invention relates to the field of information technology, in particular to a semiconductor wafer cleaning technology, and in particular to an intelligent control method for a wafer cleaning and conveying process. Background Art

[0002] Wafer cleaning is a critical step in semiconductor wafer manufacturing. Currently used wafer conveyors are prone to vibration and displacement during transport, affecting cleaning effectiveness. The wafer's position on the conveyor belt is not precise enough, leading to blind spots during cleaning and inadequate cleaning.

[0003] In addition, wafers of different sizes and weights require different conveying speeds and cleaning parameters. Existing conveying devices cannot be dynamically adjusted according to the characteristics of the wafers, which affects the efficiency and quality of cleaning.

[0004] The rotation angle and speed of the wafer during transport are also important factors affecting the cleaning effect. An inappropriate rotation angle will prevent the cleaning liquid from evenly covering the wafer surface, and too fast or too slow a rotation speed will also affect the thoroughness of the cleaning.

[0005] Therefore, it is necessary to design a wafer conveying device that can accurately control the wafer position, dynamically adjust the conveying speed, and optimize the rotation angle and speed, as well as a matching cleaning method, to improve the stability and quality of wafer cleaning and ensure the reliability and yield of semiconductor devices. Summary of the Invention

[0006] The present invention provides an intelligent control method for a wafer cleaning and conveying process, which comprises the following steps:

[0007] A machine vision system is used to detect the position coordinates of the wafer on the conveyor belt in real time. Based on the acquired coordinate data, the servo motor precisely controls the movement of the conveyor belt to adjust the wafer to the preset optimal cleaning position. The optimal cleaning position is defined as the position where the center of the wafer is aligned with the center of the cleaning nozzle. This ensures that the cleaning liquid evenly covers the wafer surface and eliminates cleaning blind spots caused by wafer position deviation.

[0008] Based on the wafer image acquired by the machine vision system, the wafer's dimensions are calculated. At the same time, a high-precision weighing sensor is used to measure the wafer's weight. A mapping table is established that maps wafer size and weight to the optimal conveying speed. This mapping table is based on experimental data and experience. The corresponding optimal conveying speed parameters are obtained from the mapping table, and the servo motor is controlled to adjust the conveyor belt speed, achieving dynamic optimization of the conveying speed.

[0009] During wafer transport, a gyroscope sensor is used to detect vibration and displacement signals from the conveyor in real time. The collected signals are input into a pre-trained support vector machine model for classification to determine whether the current vibration and displacement exceed preset thresholds. This support vector machine model is trained using historical data and can accurately distinguish between normal and abnormal vibration and displacement states.

[0010] When vibration or displacement exceeding the threshold is detected, the servo motor is triggered to perform compensation control. By adjusting the conveying speed and acceleration parameters, vibration and displacement are suppressed to ensure stable conveyance of wafers to the rotary cleaning device. The compensation control algorithm dynamically adjusts the control parameters according to the amplitude of vibration and displacement to achieve fast response and precise compensation.

[0011] For the spin cleaning process, before the cleaning fluid is sprayed, a computer vision algorithm is used to analyze images of the stationary wafer surface. Image segmentation and feature extraction are used to determine the initial state of the wafer surface. During the cleaning process, a high-speed camera is used to capture instantaneous images of the rotating wafer. Combined with the initial state analysis results, the uniformity of the cleaning fluid coverage on the wafer surface is evaluated to obtain a quantitative coverage score.

[0012] The coverage score is compared with a preset threshold. When it falls below the threshold, the wafer's rotation angle and speed are dynamically adjusted. A fuzzy control algorithm is used to optimize the rotation parameters. Based on the three input variables of coverage score, rotation angle, and speed, the algorithm outputs the optimal angle and speed adjustments. The fuzzy rules are set based on knowledge, and repeated tuning ensures the effectiveness of the algorithm. Coverage trends are continuously monitored, and cleaning parameters are optimized in real time.

[0013] Based on the coverage trend and the wafer rotation time, a reinforcement learning algorithm is used to continuously optimize the rotation angle and speed control strategy. The reinforcement learning algorithm aims to maximize cleaning thoroughness, using the coverage score as a reward signal. Through multiple iterations, the optimal rotation parameter combination is learned. The learning process considers the characteristics of different wafer types, forming a highly adaptable control strategy and improving the stability of the cleaning effect.

[0014] After each wafer is cleaned, the surface of the cleaned wafer is inspected for defects. The surface defects are identified and located through a deep learning algorithm, and a defect distribution map is generated. The defect information is correlated with the previous cleaning parameters for analysis to identify the factors that lead to incomplete cleaning. The analysis results are fed back to the upstream process parameter optimization system to adjust the cleaning fluid formula, spray pressure and other parameters, continuously improve the wafer cleaning process, and form a closed-loop optimization.

[0015] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0016] The present invention discloses an intelligent control method for the wafer cleaning and conveying process, which uses machine vision to detect the wafer position in real time, combined with high-precision weighing and vibration detection, to dynamically optimize the conveying speed and position. During the rotational cleaning stage, computer vision is used to analyze the uniformity of cleaning liquid coverage, and fuzzy control and reinforcement learning algorithms are used to adjust the rotation parameters in real time to ensure the thoroughness of cleaning. Defect detection is performed after cleaning, and the results are fed back to the process parameter optimization system to form a closed-loop improvement. The present invention realizes intelligent control of the entire wafer cleaning process, significantly improves cleaning quality and efficiency, reduces manual intervention, and provides an innovative solution for the wafer cleaning link in semiconductor manufacturing. DETAILED DESCRIPTION

[0017] To further understand the content of the present invention, the present invention is described in detail with reference to the embodiments. The present application is further described in detail with reference to the embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention.

[0018] The intelligent control method for a wafer cleaning and conveying process according to this embodiment specifically includes the following steps:

[0019] S1. A machine vision system detects the wafer's position coordinates on the conveyor belt in real time. Based on this coordinate data, a servo motor precisely controls the conveyor belt's movement, adjusting the wafer to the preset optimal cleaning position. The optimal cleaning position is defined as the point where the center of the wafer aligns with the center of the cleaning nozzle. This ensures that the cleaning fluid evenly covers the wafer surface and eliminates blind spots caused by wafer position deviation.

[0020] S1 includes: acquiring image data of the wafer on the conveyor belt, the image data being acquired by a machine vision system through real-time image acquisition; preprocessing the image data, denoising and enhancing the image data using an image preprocessing algorithm, and obtaining a preprocessed wafer image; extracting the contour features of the wafer from the preprocessed wafer image, calculating the geometric center coordinates of the wafer based on the contour features, and obtaining the real-time position coordinate data of the wafer on the conveyor belt; acquiring the preset optimal cleaning position coordinates, comparing the real-time position coordinate data of the wafer on the conveyor belt with the optimal cleaning position coordinates, and calculating the deviation between the wafer center and the optimal cleaning position; if the deviation exceeds a preset threshold, triggering the servo motor to adjust the conveyor belt movement; based on the deviation, using The PID control algorithm controls the speed and movement direction of the servo motor, and moves the wafer to the optimal cleaning position by adjusting the movement of the conveyor belt so that the center of the wafer is aligned with the center of the cleaning nozzle; when the wafer reaches the optimal cleaning position, the cleaning nozzle is triggered to start, the flow rate and spraying time of the cleaning liquid are controlled, and the cleaning liquid is evenly sprayed on the center of the wafer so that the cleaning liquid completely covers the wafer surface; after cleaning is completed, the machine vision system is used to capture and analyze images of the wafer surface to detect the cleanliness of the wafer surface; if the cleanliness does not meet the preset standard, the cleaning process is triggered again until the wafer surface is completely clean; the qualified wafers are transferred to the next process, the cleaning process data are recorded and statistically analyzed, the cleaning effect is evaluated, and the cleaning parameters are dynamically optimized.

[0021] Specifically, machine vision systems are a key technology in the wafer cleaning process. They use high-resolution cameras to capture real-time images of wafers on a conveyor belt, determining their position and surface condition. For example, a 10-megapixel industrial camera can capture images at 30 frames per second, ensuring that even the finest features on the wafer are captured. Image preprocessing is a crucial step in improving wafer image quality. Common preprocessing methods include Gaussian filtering for denoising and histogram equalization for contrast enhancement. Gaussian filtering effectively removes random noise from the image, while histogram equalization enhances overall image contrast, making wafer edges more distinct. Edge detection algorithms are used to extract wafer contour features. The Canny algorithm is a commonly used edge detection method. It calculates the gradient magnitude and direction of the image and performs non-maximum suppression and dual-threshold detection to accurately extract the wafer edge contour. Based on the extracted contour, the geometric center coordinates of the wafer can be calculated, thereby determining the wafer's real-time position on the conveyor belt. PID control algorithms are used to precisely control the servo motor and adjust the conveyor belt's motion. The PID controller calculates the appropriate control signal based on the deviation between the wafer center and the optimal cleaning position. For example, if the deviation is 5mm, the PID controller might output a corresponding voltage signal, causing the servo motor to operate at the appropriate speed and direction to move the wafer to the optimal cleaning position. Controlling the cleaning nozzle is key to ensuring uniform cleaning across the wafer surface. A piezoelectric nozzle can be used to control the flow rate and spray range of the cleaning fluid by adjusting the vibration frequency of the piezoelectric crystal. For example, the nozzle can be set to vibrate at 300Hz, spraying 0.5ml of cleaning fluid per second for 10 seconds to ensure complete coverage of the wafer surface. Cleanliness inspection is a key step in evaluating cleaning effectiveness. Image analysis techniques, such as grayscale threshold segmentation or texture analysis, can be used to detect residue on the wafer surface. For example, a cleanliness threshold of 99.9% of the surface area free of visible residue can be set as a qualified cleaning standard. Data logging and statistical analysis help continuously optimize the cleaning process. Data such as cleaning parameters, cleaning time, and cleanliness can be recorded for each wafer. Big data analysis techniques, such as machine learning algorithms, can be used to identify the relationship between cleaning performance and these parameters, allowing dynamic adjustment of cleaning parameters to improve cleaning efficiency and quality. This data-driven optimization approach can significantly improve the overall efficiency of wafer cleaning, reduce unnecessary re-cleaning, and save time and resources.

[0022] S2. Calculate wafer dimensions based on wafer images captured by the machine vision system. Simultaneously, use a high-precision load cell to measure wafer weight. Create a mapping table based on experimental data and experience, linking wafer size and weight to optimal conveyor speed. Query the table to obtain the optimal conveyor speed parameters, and control the servo motor to adjust the conveyor belt speed, achieving dynamic conveyor speed optimization.

[0023] S2 includes: acquiring a wafer image, and using an image processing algorithm to determine the diameter and thickness information of the wafer in the wafer image; acquiring wafer weight measurement data collected by a high-precision weighing sensor to obtain the weight information of the wafer; according to the diameter, thickness and weight information of the wafer, querying in a pre-constructed wafer parameter and conveying speed mapping table to obtain the optimal conveying speed parameters corresponding to the current wafer parameters; if the optimal conveying speed parameters are different from the current conveyor belt speed, controlling the servo motor to adjust the conveyor belt speed to the optimal conveying speed parameters; during the wafer conveying process, continuously acquiring wafer images and weighing sensor data, updating the optimal conveying speed parameters in real time, and realizing dynamic optimization of the conveying speed; using a machine learning algorithm, based on historical wafer parameters and conveying speed data, continuously optimizing the wafer parameters and conveying speed mapping table to improve the accuracy of conveying speed optimization; converting wafer production experience into optimization rules, supplementing and improving the wafer parameters and conveying speed mapping table, to meet the conveying speed optimization under special process requirements.

[0024] Specifically, wafer image processing is a key step in optimizing conveyor speed. Wafer images are captured using a high-resolution camera, and edge detection algorithms are used to extract the wafer outline. Methods such as the Hough transform are then applied to calculate the wafer diameter. For example, an 8-inch wafer might measure a diameter of 200.3 mm. Laser triangulation is used to measure thickness, achieving micron-level accuracy. A typical 8-inch wafer is approximately 725 microns thick. High-precision load cells are crucial for obtaining wafer weight information. Strain gauge load cells offer an accuracy of 0.01 gram. For example, an 8-inch wafer typically weighs approximately 53 grams. Combining size and weight information allows inferring the wafer's material and processing status, which is crucial for determining the optimal conveyor speed. A pre-built mapping table bridges wafer parameters and conveyor speed. The table might include the following: an 8-inch silicon wafer with a thickness of 725 microns and a weight of 53 grams corresponds to an optimal conveyor speed of 0.5 m / s. If the detected wafer parameters match an entry in the table, the system will use the corresponding speed parameters. Servo motor control is the core of achieving precise speed regulation. A closed-loop control system uses encoder feedback to provide real-time speed information, which is then compared with the target speed for PID control. For example, if the current speed is 0.4 m / s and the target speed is 0.5 m / s, the system will gradually increase the motor speed until the target value is reached. Dynamic optimization is key to improving system adaptability. During the conveying process, the system continuously collects wafer data. If minor defects are detected on the wafer surface, the conveying speed may be appropriately reduced to ensure safety. For example, if a microcrack is detected on the wafer edge, the system may reduce the speed from 0.5 m / s to 0.4 m / s. The application of machine learning algorithms enables the system to self-optimize. Using algorithms such as support vector machines (SVMs), the system can learn the complex relationship between wafer parameters and optimal conveying speed. For example, the system may discover that the optimal conveying speed for a certain type of wafer varies slightly under specific temperature and humidity conditions and update the mapping table accordingly. This system addresses the shortcomings of purely data-driven approaches. By encoding engineers' experience into a rule base, the system can handle special cases. For example, for wafers that have just completed a polishing process, the system may recommend reducing the conveyor speed by 20% to prevent possible surface scratches. This approach is particularly useful for processing new or small-batch wafers, where historical data may not be sufficient for machine learning algorithms to make accurate decisions.

[0025] During wafer transport, a gyroscope sensor is used to detect vibration and displacement signals from the conveyor in real time. These signals are then fed into a pre-trained support vector machine model for classification, determining whether the vibration and displacement exceed preset thresholds. This support vector machine model, trained using historical data, accurately distinguishes between normal and abnormal vibration and displacement states.

[0026] S3 includes: acquiring real-time vibration signals and displacement signals of the wafer conveying device, wherein the vibration signals and displacement signals are collected by a gyroscope sensor provided on the conveying device; preprocessing the vibration signals and displacement signals to obtain standardized feature data, wherein the preprocessing includes denoising and normalization; inputting the standardized feature data into a pre-trained support vector machine model, and using the support vector machine model to classify and judge the standardized feature data to determine whether the vibration signals and displacement signals are within a normal range; if the result of the classification judgment shows that the vibration signals and displacement signals exceed a preset threshold value, it is determined that the current state is abnormal and an early warning signal is output; if the result of the classification judgment shows that the vibration signals and displacement signals do not exceed the preset threshold value, it is determined that the current state is normal, and the vibration signals and displacement signals at the next moment are continuously acquired and monitored; continuously recording the vibration signals, displacement signals and corresponding classification judgment results at each moment to form a historical data set; regularly using the historical data set to retrain the support vector machine model to obtain an optimized support vector machine model for subsequent classification judgment.

[0027] Specifically, gyroscope sensors play a crucial role in the wafer conveying process, capturing vibration and displacement information of the conveyor in real time. For example, a high-precision gyroscope can detect angular velocity changes as small as 0.01 degrees per second, which is crucial for monitoring subtle vibrations during wafer conveyance. The acquired raw signals often contain noise and require filtering. Common filtering methods include Kalman filtering and wavelet transform, which effectively remove high-frequency noise while preserving useful signal features. During data preprocessing, normalization is used to unify data of varying dimensions to the same scale. For example, min-max normalization can be used to map both vibration amplitude and displacement to a range of 0 to 1. This processed data is more suitable for input into machine learning models. Support vector machine (SVM) models excel at anomaly detection, finding the optimal classification hyperplane in high-dimensional feature spaces. In wafer conveyance monitoring, radial basis function (RBF) kernels can be used. These kernels are capable of handling nonlinear classification problems and are suitable for capturing complex patterns in vibration and displacement data. The design of early warning systems requires a trade-off between sensitivity and false alarm rate. For example, multiple warning thresholds can be set: a mild warning is issued when the vibration amplitude exceeds 1.5 times the normal range, and a severe warning is issued when it exceeds 2 times. This tiered warning mechanism helps engineers better assess equipment status and take timely maintenance measures. Continuously recorded historical data is used not only for model training but also for equipment lifespan prediction. By analyzing long-term trends in vibration and displacement data, it is possible to predict when equipment may need major repairs or replacement, thereby implementing preventive maintenance. Regular model retraining is key to maintaining system performance. An incremental learning approach can be employed, with the model updated every time a certain amount of new data is accumulated (e.g., weekly or monthly). This approach allows the model to adapt to subtle changes in the equipment over time, such as changes in vibration characteristics due to component wear. In practical applications, it may also be necessary to consider differences in vibration characteristics under different process conditions. For example, the normal vibration range may vary when processing wafers of different sizes or materials. Therefore, multiple SVM models can be established for different process conditions, with the model automatically switching based on the current production task to improve the accuracy of anomaly detection. Furthermore, integrating other sensor data can further enhance the reliability of the monitoring system. For example, it can simultaneously monitor parameters such as motor current and temperature, integrating multiple data sources for anomaly detection. This multimodal data fusion approach provides more comprehensive equipment status information, reduces false alarms, and improves the accuracy and reliability of early warnings. This comprehensive and sophisticated monitoring and early warning system can significantly improve the stability and reliability of the wafer transport process, reduce product losses and downtime caused by equipment failures, and thus improve overall production efficiency.

[0028] S4. When vibration or displacement exceeding a threshold is detected, the servo motor is triggered to perform compensation control. By adjusting the conveying speed and acceleration parameters, vibration and displacement are suppressed, ensuring stable wafer delivery to the spin cleaning device. The compensation control algorithm dynamically adjusts control parameters based on the amplitude of vibration and displacement, achieving rapid response and precise compensation.

[0029] S4 includes: acquiring real-time vibration and displacement data during wafer conveying, the real-time vibration and displacement data being obtained by collecting vibration and displacement signals through sensors, and converting the vibration and displacement signals into digital quantities; determining whether the real-time vibration and displacement data exceed a preset vibration threshold and displacement threshold range; if the real-time vibration and displacement data exceed the preset vibration threshold and displacement threshold range, triggering a servo motor to perform compensation control; filtering the collected vibration and displacement data using a Kalman filter algorithm to obtain filtered vibration and displacement data; calculating compensation control parameters of the servo motor using a PID control algorithm based on the filtered vibration and displacement data, the compensation control parameters including adjustment values ​​for conveying speed and acceleration; transmitting the compensation control parameters to a servo motor controller, and achieving vibration and displacement suppression during wafer conveying by adjusting the speed and acceleration of the servo motor; during the compensation control process, continuously monitoring the vibration and displacement status of the wafer, and dynamically optimizing the compensation control parameters using an adaptive control algorithm; stopping compensation control when the wafer is stably conveyed to the rotary cleaning device, and recording relevant data during the compensation control process.

[0030] Specifically, controlling vibration and displacement during wafer transport is a critical step in ensuring wafer quality. Real-time data acquisition typically utilizes high-precision accelerometers and displacement sensors, such as MEMS accelerometers and laser displacement sensors. These sensors can capture minute vibration and displacement changes with micron-level accuracy. For example, the accelerometers used in one wafer production line have a sensitivity of 0.1g, the displacement sensor has a resolution of 0.1μm, and a sampling frequency of 1kHz, enabling real-time monitoring of subtle changes during transport. Threshold setting is crucial for detecting abnormal conditions. These thresholds are typically determined based on wafer material properties and process requirements. For example, a 300mm wafer production line sets a vibration threshold of 0.5g and a displacement threshold of ±50μm. When these thresholds are exceeded, the system immediately triggers compensation control to prevent wafer damage. Kalman filtering algorithms play a crucial role in signal processing, effectively removing random noise and improving data reliability. For example, after applying Kalman filtering at one wafer fab, the signal-to-noise ratio of vibration signals increased by 30% and the accuracy of displacement data by 20%, significantly improving the accuracy of subsequent control. PID control algorithms are central to achieving precise compensation. By adjusting the proportional, integral, and differential parameters, vibration and displacement can be quickly responded to and stably suppressed. In practical applications, a wafer conveyor system uses adaptive PID control to automatically adjust parameters based on different operating conditions. For example, during the startup phase, the proportional coefficient is increased for faster response; during the stabilization phase, the integral coefficient is increased to eliminate static errors; and during sudden disturbances, the differential coefficient is increased to improve system stability. Servo motor control is a key component in implementing compensation. High-precision servo motors can achieve micron-level position adjustment and millisecond-level response speeds. For example, a wafer conveyor system uses a servo motor with a resolution of 0.01°, coupled with a reducer to achieve a minimum adjustment of 0.1μm, accurately executing the control commands output by the PID algorithm. Adaptive control algorithms dynamically optimize control parameters based on real-time monitoring data. For example, a wafer fab uses a fuzzy neural network adaptive algorithm to continuously adjust control strategies based on historical data and real-time feedback. After one month of operation, this system achieved a 15% improvement in vibration suppression and a 10% increase in displacement control accuracy. Data recording and analysis are crucial for system optimization. For example, a wafer production line has established a comprehensive database that records each batch of wafer conveying parameters, vibration displacement data, and control effectiveness. Through big data analysis, they discovered a correlation between equipment aging and increased vibration, enabling proactive maintenance forecasts and reducing unplanned downtime by 30%. The combined application of these technologies not only ensures the stability of the wafer conveying process but also improves production efficiency and yield. After implementing this system at a large wafer fab, the wafer breakage rate was reduced by 50%, product yield increased by 3 percentage points, and annual cost savings reached tens of millions of yuan.

[0031] S5. For the spin cleaning process, before the cleaning fluid is sprayed, a computer vision algorithm analyzes images of the stationary wafer surface. Image segmentation and feature extraction are used to determine the initial state of the wafer surface. During the cleaning process, a high-speed camera captures instantaneous images of the rotating wafer. Combined with the initial state analysis results, the uniformity of the cleaning fluid coverage on the wafer surface is evaluated to obtain a quantitative coverage score.

[0032] S5 involves capturing stationary wafer surface images using a high-resolution camera. Image preprocessing algorithms then convert the images to grayscale to enhance image contrast, providing high-quality input for subsequent image segmentation and feature extraction. Edge detection algorithms are used to segment the preprocessed images, accurately extracting wafer edge contours. A region growing algorithm is then used to segment the wafer surface regions and obtain feature data of the wafer's initial state. Feature information for surface defects and stains is extracted from the segmented images, and a feature matching algorithm is used to compare these features with a pre-defined defect database to determine the initial classification label for the wafer surface. During the cleaning process, a high-speed camera captures instantaneous images of the rotating wafer in real time. An image registration algorithm is used to align the instantaneous images with the initial state images to ensure spatiotemporal consistency in image analysis. The registered instantaneous images are converted to color space to extract the color characteristics of the cleaning solution on the wafer surface. A regional coverage analysis algorithm is then used to assess the cleaning solution's coverage and uniformity. Based on the coverage and uniformity analysis results, a quantitative coverage score is calculated using a pre-defined coverage scoring model that considers metrics such as coverage area ratio and distribution uniformity. If the coverage score is lower than the preset threshold, feedback is provided to adjust the cleaning liquid injection parameters and the cleaning process is repeated until the score reaches the preset standard to ensure that the wafer surface cleaning effect meets the process requirements.

[0033] Specifically, capturing wafer surface images with a high-resolution camera is fundamental to cleaning quality assessment. Industrial cameras with 4K or 8K resolution, combined with high-quality lenses, are typically used to capture minute details on the wafer surface. Grayscaling, part of image preprocessing, converts RGB color images into single-channel grayscale images, simplifying subsequent processing steps. Histogram equalization is commonly used for contrast enhancement, redistributing pixel values ​​within the 0-255 range to improve overall image contrast. Edge detection is a key step in image segmentation. The Canny algorithm is a commonly used edge detection method, using Gaussian filtering, non-maximum suppression, and double-threshold detection to accurately extract wafer edges. The region growing algorithm, starting from a preset seed point, gradually expands regions of similar pixels, ultimately capturing a complete region of the wafer surface. Feature extraction and matching are key to identifying wafer surface conditions. Extractable features include texture features (such as the gray-level co-occurrence matrix) and shape features (such as area, perimeter, and roundness). Feature matching can utilize the KNN (K-nearest neighbor) algorithm, comparing extracted features against a pre-established defect library to determine the initial wafer surface condition classification. High-speed cameras capture images of the rotating wafer in real time, typically at acquisition rates exceeding 1000 frames per second, to ensure clear instantaneous images. Image registration algorithms such as SIFT (Scale-Invariant Feature Transform) are used to align the instantaneous image with the initial state image, compensating for differences caused by rotation and displacement. Color space conversion typically involves RGB to HSV, which is more suitable for describing the color characteristics of the cleaning fluid. Area coverage analysis uses HSV thresholds to extract the cleaning fluid coverage area and calculate the coverage ratio and distribution uniformity. A coverage scoring model can comprehensively evaluate cleaning performance by considering coverage ratio (e.g., greater than 95%) and uniformity indicators (e.g., standard deviation less than 0.1). If the coverage score falls short of the target, the system automatically adjusts the cleaning fluid spray parameters. For example, increasing the spray pressure (e.g., from 2 bar to 2.5 bar), adjusting the nozzle angle (e.g., from 60° to 75°), or extending the cleaning time (e.g., from 30 seconds to 35 seconds). This closed-loop control ensures adaptive optimization of the cleaning process, ultimately achieving the preset cleaning quality standard. The entire process demonstrates the application of computer vision technology in semiconductor manufacturing. Through real-time image analysis and feedback control, the accuracy and efficiency of wafer cleaning are significantly improved. This not only reduces the subjectivity and labor intensity of manual inspection, but also enables real-time monitoring and optimization of the cleaning process, which is of great significance for improving product yield and production efficiency.

[0034] S6 compares the coverage score to a preset threshold. If it falls below the threshold, the wafer's rotation angle and speed are dynamically adjusted. Rotation parameters are optimized using a fuzzy control algorithm. Based on the three input variables of coverage score, rotation angle, and speed, the algorithm outputs the optimal angle and speed adjustments. Fuzzy rules are defined based on knowledge, and algorithmic effectiveness is ensured through repeated optimization. Coverage trends are continuously monitored, allowing for real-time optimization of cleaning parameters.

[0035] S6 includes: obtaining real-time coverage score data, comparing the real-time coverage score data with a preset threshold, and triggering a fuzzy control algorithm if the real-time coverage score data is lower than the preset threshold; inputting the current coverage score, rotation angle and speed into the fuzzy control algorithm, calculating according to preset fuzzy rules, and obtaining an optimal angle adjustment amount and an optimal speed adjustment amount; feeding back the optimal angle adjustment amount and the optimal speed adjustment amount to the cleaning equipment, and dynamically adjusting the wafer rotation parameters; continuously monitoring coverage changes, and if the adjusted coverage score is still lower than the preset threshold, restarting the fuzzy control algorithm for optimization; recording the coverage score and the corresponding rotation angle and speed parameters after each adjustment, and constructing an optimized data set; iteratively improving the fuzzy rules in the fuzzy control algorithm according to the optimized data set to obtain an improved fuzzy control algorithm; using the improved fuzzy control algorithm to optimize and control the subsequent wafer cleaning process, thereby improving the adaptability and accuracy of the cleaning effect.

[0036] Specifically, obtaining real-time coverage score data is a critical step in the cleaning process. For example, suppose the preset coverage threshold is 90%, and the real-time monitoring system indicates the current coverage is 85%. At this point, the system immediately triggers a fuzzy control algorithm to ensure that cleaning results meet the target. The fuzzy control algorithm's inputs include the current coverage score, rotation angle, and speed. For example, assume the current rotation angle is 30 degrees and the speed is 100 revolutions per minute. Based on pre-set fuzzy rules, the algorithm performs the following calculations: If the coverage is low and the rotation speed is fast, the speed is appropriately reduced; if the coverage is low and the rotation angle is small, the angle is appropriately increased. Through fuzzy logic reasoning, the algorithm outputs the optimal angle adjustment of 5 degrees and the speed adjustment of 10 revolutions per minute. These adjustments are fed back to the cleaning equipment, which dynamically adjusts the wafer rotation parameters to 35 degrees and 90 revolutions per minute. Coverage changes are continuously monitored. If the score remains below the threshold, the system reactivates the fuzzy control algorithm for optimization. The coverage score and parameter changes after each adjustment are recorded to construct an optimized dataset. For example, after the first adjustment, the coverage was 88%, and after the second adjustment, it was 92%. This data is used to iteratively improve the algorithm. Through multiple iterations, the accuracy and adaptability of the fuzzy rules are enhanced. Improving the accuracy of the fuzzy rules means the algorithm can more accurately predict the impact of adjustments on coverage. For example, an initial algorithm may have overestimated the impact of speed adjustments on coverage. After iteration, the algorithm can more accurately balance speed and angle adjustments, ensuring that each adjustment effectively improves coverage. Improved adaptability is reflected in the algorithm's ability to adapt to different cleaning conditions. For example, different wafer types may respond differently to cleaning fluids. Through multiple iterations, the algorithm can adjust the fuzzy rules based on the characteristics of each wafer, ensuring optimal cleaning results under various conditions. This not only improves cleaning efficiency but also reduces cleaning fluid waste and lowers production costs. For example, through optimization, cleaning fluid usage was reduced by 10% and cleaning time was shortened by 5%, significantly improving overall production line efficiency. Furthermore, recording and analyzing the data from each adjustment helps identify potential issues and areas for improvement. For example, the data may reveal that cleaning results are consistently suboptimal at certain angles and speeds. In-depth analysis can identify the cause and implement targeted improvements. Continuously optimizing cleaning results also improves product quality. For example, if tiny defects and stains on the wafer surface are not thoroughly cleaned, they can amplify defects in subsequent processes and affect the performance of the final product. By precisely controlling the cleaning process, we ensure a highly clean wafer surface, thereby improving product yield and reliability. In short, through real-time monitoring, dynamic adjustment of fuzzy control algorithms, and data-driven iterative optimization, the cleaning process is continuously improved, ensuring that the wafer surface cleaning effect meets high process standards, thereby improving production efficiency and product quality.

[0037] S7. Based on coverage trends and wafer rotation time, a reinforcement learning algorithm is used to continuously optimize the rotation angle and speed control strategy. Aiming to maximize cleaning thoroughness, the reinforcement learning algorithm uses coverage scores as reward signals and learns the optimal rotation parameter combination through multiple iterations. The learning process considers the characteristics of different wafer types, forming a highly adaptable control strategy and improving the stability of cleaning results.

[0038] The system obtains wafer type information and, based on a pre-established database of wafer characteristics, determines the initial control parameters for the spin cleaning process for that type of wafer, including the rotation angle range and speed range. The wafer is placed in the cleaning equipment, the cleaning process is initiated, and the rotation angle and speed are set according to the current control parameters. The wafer is then spin-cleaned. A visual sensor captures real-time images of the wafer surface, processes the images, and calculates a coverage curve for the wafer surface during the cleaning process. This coverage curve is compared with a preset ideal coverage curve, and the similarity between the two curves is calculated. This similarity value serves as an immediate reward for the reinforcement learning algorithm. Based on the reward, the reinforcement learning algorithm updates the parameters of the control policy network using a policy gradient method to obtain new rotation angle and speed control values. The updated control parameters are applied to the wafer cleaning equipment, adjusting the rotation angle and speed. Steps 3-5 are repeated for multiple rounds of iterative learning optimization. When the similarity between the coverage curve and the ideal curve reaches a preset threshold, or the number of iterations reaches an upper limit, the learning process is terminated, and the optimized rotation angle and speed control strategy is output to guide subsequent cleaning processes for wafers of the same type.

[0039] Specifically, a wafer characteristic database is established based on information such as the physical and chemical properties and surface structure of different wafer types. For example, for an 8-inch silicon wafer, the initial rotation angle range may be set to 0-180 degrees, and the speed range may be 500-2000 rpm. These parameters are selected based on factors such as wafer size, material strength, and cleaning requirements. When capturing images of the wafer surface, a vision sensor can use a high-resolution CCD camera in conjunction with an appropriate lighting system. The image processing process includes preprocessing steps such as denoising and contrast enhancement. Cleaned areas are then identified through algorithms such as edge detection and region segmentation. Coverage can be calculated using pixel statistics, such as dividing the number of pixels in the cleaned area by the total number of pixels. An ideal coverage curve typically exhibits an S-shape, with slow initial growth, rapid mid-term increases, and a flattening trend in the later stages. The similarity between the actual and ideal curves can be quantified by calculating the Euclidean distance or correlation coefficient between the two curves. For example, if the correlation coefficient between the two curves is 0.95, 0.95 can be input as the immediate reward value into a reinforcement learning algorithm. Reinforcement learning algorithms can employ deep Q-networks (DQNs) or policy gradient methods. Taking the DQN as an example, the input state may include the current coverage, rotation angle, and speed, and the output is the Q value for different angle and speed combinations. The ε-greedy strategy selects actions and gradually optimizes the Q network parameters. For example, the initial ε value is set to 0.9 and gradually reduced to 0.1 as training progresses to balance exploration and exploitation. During the control parameter update process, a gradual adjustment strategy can be adopted. For example, if the current angle is 90 degrees and the speed is 1000 rpm, and the DQN outputs the optimal action to increase the angle and decrease the speed, the angle can be adjusted to 100 degrees and the speed to 950 rpm. This gradual adjustment helps maintain the stability of the cleaning process. During the iterative learning optimization process, appropriate stopping conditions must be set. For example, the similarity threshold can be set to 0.98 and the maximum number of iterations to 100. Optimization stops when any of these conditions are met. The optimization results can be represented as a policy network that takes as input the wafer type and cleaning stage and outputs the optimal angle and speed control values. This reinforcement learning-based optimization method is more adaptable to variations in wafer types and cleaning environments than traditional fixed-parameter control. Through continuous learning and optimization, the system can gradually improve cleaning efficiency and quality, reduce manual intervention, and achieve intelligent control. At the same time, the accumulated optimization experience can be used to guide the setting of cleaning parameters for new wafers, improving overall production efficiency.

[0040] S8. After each wafer is cleaned, the surface of the cleaned wafer is inspected for defects. Surface defects, such as contamination and scratches, are identified and located using a deep learning algorithm, generating a defect distribution map. Defect information is correlated with previous cleaning parameters and analyzed to identify factors contributing to incomplete cleaning. The analysis results are fed back to the upstream process parameter optimization system, which adjusts parameters such as the cleaning fluid formulation and spray pressure to continuously improve the wafer cleaning process, forming a closed-loop optimization process.

[0041] Acquire an image of the wafer surface after cleaning, the image being a defect image; preprocess the defect image to obtain a preprocessed image, the preprocessing including denoising and contrast enhancement; use a convolutional neural network to perform defect recognition on the preprocessed image to obtain a defect distribution map; acquire cleaning parameter data, the cleaning parameter data including a cleaning fluid formula and a spray pressure, and construct a parameter database; perform correlation analysis on the defect distribution map and the parameter database, use a random forest algorithm to analyze the correlation between defects and cleaning parameters, and obtain correlation analysis results; determine the cleaning parameters that need to be adjusted based on the correlation analysis results, and generate optimization suggestions; transmit the optimization suggestions to a process parameter optimization system to adjust the cleaning fluid formula and the spray pressure; monitor the cleaning effect in real time through a closed-loop control system, provide feedback on the adjustment results, and continuously optimize the cleaning process.

[0042] Specifically, deep learning algorithms play a vital role in wafer surface defect detection. For example, convolutional neural networks, through multi-layer convolution and pooling operations, extract image features and effectively identify defects such as tiny scratches and particle contamination. In practical applications, advanced network structures such as ResNet can be used to improve detection accuracy. For example, a semiconductor manufacturer employed a modified YOLOv5 model, increasing its defect detection rate to 99.5%, significantly surpassing traditional machine vision methods. Image preprocessing is crucial for improving defect detection. Gaussian filtering effectively removes image noise, while histogram equalization enhances image contrast. In one case, the introduction of an adaptive threshold segmentation algorithm successfully addressed the problem of reflective interference on the wafer surface, making defect edges more clearly discernible. Generating defect distribution maps lays the foundation for subsequent analysis. Heat maps can visually display areas with high defect concentrations, helping engineers quickly locate problems. Using this method, one manufacturer discovered elevated defect rates in edge areas and subsequently optimized the cleaning nozzle layout, significantly improving cleaning uniformity. The construction of a cleaning parameter database facilitates comprehensive analysis of process influencing factors. In addition to cleaning fluid formulation and spray pressure, key parameters such as temperature and time should also be considered. A study has shown that precisely controlling the pH of the cleaning fluid can increase particle removal rates by 15%. Machine learning algorithms play a key role in parameter optimization. The random forest algorithm has excellent feature importance assessment capabilities, allowing it to quickly identify key factors influencing defect generation. In one study, random forest analysis revealed a strong correlation between cleaning fluid concentration and micro-scratch defects. Adjusting the formulation accordingly reduced the defect rate by 30%. A closed-loop control system ensures continuous process optimization. Real-time monitoring of cleaning results allows for timely parameter adjustments to avoid batch-to-batch quality fluctuations. A production line has introduced an adaptive system based on fuzzy control, dynamically adjusting cleaning pressure based on defect detection results, maintaining a stable yield rate above 99%. This data-driven cleaning process optimization approach not only improves product quality but also reduces production costs. By reducing chemical consumption and increasing equipment utilization, one factory has achieved annual cost savings exceeding 1 million yuan. Furthermore, this method's high degree of automation reduces human intervention, improves process consistency and repeatability, and provides strong support for the intelligent transformation of the semiconductor manufacturing industry.

[0043] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent control method for a wafer cleaning and conveying process, characterized in that: The method comprises the following steps: S1. Use a machine vision system to detect the position coordinates of the wafer on the conveyor belt in real time. Based on the acquired coordinate data, the servo motor precisely controls the movement of the conveyor belt to adjust the wafer to the preset optimal cleaning position. The optimal cleaning position is defined as the position where the center of the wafer is aligned with the center of the cleaning nozzle. This ensures that the cleaning liquid evenly covers the wafer surface and eliminates cleaning blind spots caused by wafer position deviation. S2. Calculate wafer dimensions based on wafer images captured by the machine vision system. Simultaneously, use a high-precision load cell to measure wafer weight and establish a mapping table between wafer size and weight and optimal conveying speed. This mapping table is based on experimental data and experience. Query the corresponding optimal conveying speed parameters from the mapping table and control the servo motor to adjust the conveyor belt speed, achieving dynamic optimization of the conveying speed. S3. During wafer transport, a gyroscope sensor is used to detect vibration and displacement signals of the transport mechanism in real time. The collected signals are input into a pre-trained support vector machine model for classification to determine whether the current vibration and displacement exceed preset thresholds. This support vector machine model is trained using historical data and can accurately distinguish between normal and abnormal vibration and displacement states. S4. When vibration or displacement exceeding a threshold is detected, the servo motor is triggered to perform compensation control. By adjusting the conveying speed and acceleration parameters, vibration and displacement are suppressed to ensure stable conveyance of the wafer to the rotary cleaning device. The compensation control algorithm dynamically adjusts the control parameters according to the amplitude of vibration and displacement to achieve fast response and precise compensation. S5. For the spin cleaning process, before the cleaning fluid is sprayed, a computer vision algorithm is used to analyze the static wafer surface image. Image segmentation and feature extraction are used to determine the initial state of the wafer surface. During the cleaning process, a high-speed camera is used to capture instantaneous images of the rotating wafer. Combined with the initial state analysis results, the uniformity of the cleaning fluid coverage on the wafer surface is evaluated to obtain a quantitative coverage score. S6. Compare the coverage score with a preset threshold. If it is below the threshold, dynamically adjust the wafer rotation angle and speed. Optimize the rotation parameters using a fuzzy control algorithm. The algorithm outputs the optimal angle and speed adjustment based on three input variables: coverage score, rotation angle, and speed. The fuzzy rules are set based on knowledge. Repeated tuning ensures the effectiveness of the algorithm, continuously monitors coverage trends, and optimizes cleaning parameters in real time. S7. Based on the coverage trend and the wafer rotation time, a reinforcement learning algorithm is used to continuously optimize the rotation angle and speed control strategy. The reinforcement learning algorithm aims to maximize cleaning thoroughness and uses the coverage score as a reward signal. The optimal rotation parameter combination is learned through multiple iterations. The characteristics of different wafer types are considered during the learning process, forming a highly adaptable control strategy and improving the stability of the cleaning effect. S8. After each wafer is cleaned, defect detection is performed on the cleaned wafer surface. Surface defects are identified and located through deep learning algorithms, and a defect distribution map is generated. The defect information is correlated with the previous cleaning parameters for analysis to identify factors leading to incomplete cleaning. The analysis results are fed back to the upstream process parameter optimization system to continuously improve the wafer cleaning process and form a closed-loop optimization.

2. The intelligent control method for wafer cleaning and conveying process according to claim 1, characterized in that: The S1 includes: Acquire image data of wafers on a conveyor belt, the image data being acquired by a machine vision system through real-time image acquisition; Preprocessing the image data, denoising and enhancing the image data using an image preprocessing algorithm to obtain a preprocessed wafer image; Extracting wafer contour features from the pre-processed wafer image, calculating the geometric center coordinates of the wafer based on the contour features, and obtaining real-time position coordinate data of the wafer on the conveyor belt; Obtaining preset optimal cleaning position coordinates, comparing the real-time position coordinate data of the wafer on the conveyor belt with the optimal cleaning position coordinates, and calculating the deviation between the center of the wafer and the optimal cleaning position; If the deviation exceeds a preset threshold, the servo motor is triggered to adjust the conveyor belt movement; According to the deviation, a PID control algorithm is used to control the speed and movement direction of the servo motor, and the wafer is moved to the optimal cleaning position by adjusting the movement of the conveyor belt so that the center of the wafer is aligned with the center of the cleaning nozzle; When the wafer reaches the optimal cleaning position, the cleaning nozzle is triggered to start, the flow rate and spraying time of the cleaning liquid are controlled, and the cleaning liquid is evenly sprayed at the center of the wafer so that the cleaning liquid completely covers the wafer surface; After cleaning is completed, the machine vision system is used to collect and analyze images of the wafer surface to detect the cleanliness of the wafer surface; If the cleanliness does not meet the preset standard, the cleaning process is triggered again until the wafer surface is completely clean; The wafers that pass the cleaning process are transferred to the next process, the cleaning process data is recorded and statistically analyzed, the cleaning effect is evaluated, and the cleaning parameters are dynamically optimized.

3. The intelligent control method for wafer cleaning and conveying process according to claim 1, characterized in that: The S2 includes: Acquire a wafer image, and determine the diameter and thickness information of the wafer in the wafer image using an image processing algorithm; Obtaining wafer weight measurement data collected by a high-precision weighing sensor to obtain weight information of the wafer; According to the diameter, thickness and weight information of the wafer, a pre-built wafer parameter and conveying speed mapping table is searched to obtain the optimal conveying speed parameters corresponding to the current wafer parameters; If the optimal conveying speed parameter is different from the current conveyor belt speed, controlling the servo motor to adjust the conveyor belt speed to the optimal conveying speed parameter; During the wafer conveying process, wafer images and weighing sensor data are continuously acquired, and the optimal conveying speed parameters are updated in real time to achieve dynamic optimization of the conveying speed; Using a machine learning algorithm, based on historical wafer parameter and conveyor speed data, the wafer parameter and conveyor speed mapping table is continuously optimized to improve the accuracy of conveyor speed optimization; The wafer production experience is converted into optimization rules, and the wafer parameter and conveying speed mapping table is supplemented and improved to meet the conveying speed optimization under special process requirements.

4. The intelligent control method for a wafer cleaning and conveying process according to any one of claims 1 to 3, characterized in that: The S3 includes: Acquiring real-time vibration signals and displacement signals of a wafer conveyor device, wherein the vibration signals and displacement signals are collected by a gyroscope sensor provided on the conveyor device; Preprocessing the vibration signal and the displacement signal to obtain standardized feature data, wherein the preprocessing includes denoising and normalization; Inputting the standardized feature data into a pre-trained support vector machine model, and using the support vector machine model to classify and judge the standardized feature data to determine whether the vibration signal and displacement signal are within a normal range; If the classification result indicates that the vibration signal and the displacement signal exceed the preset threshold, it is determined that the current state is abnormal and an early warning signal is output; If the result of the classification judgment indicates that the vibration signal and the displacement signal do not exceed the preset threshold, it is determined that the current state is normal, and the vibration signal and the displacement signal at the next moment are continuously acquired and monitored; Continuously record the vibration signal and displacement signal at each moment and the corresponding classification and judgment results to form a historical data set; The support vector machine model is retrained regularly using the historical data set to obtain an optimized support vector machine model for subsequent classification judgment.

5. The intelligent control method for a wafer cleaning and conveying process according to any one of claims 1 to 3, characterized in that: The S4 includes: Acquiring real-time vibration and displacement data during wafer transport, wherein the real-time vibration and displacement data are obtained by collecting vibration and displacement signals through sensors and converting the vibration and displacement signals into digital quantities; Determining whether the real-time vibration and displacement data exceed a preset vibration threshold and displacement threshold range; If the real-time vibration and displacement data exceed the preset vibration threshold and displacement threshold range, the servo motor is triggered to perform compensation control; Using a Kalman filter algorithm to filter the collected vibration and displacement data to obtain filtered vibration and displacement data; Calculating compensation control parameters of the servo motor using a PID control algorithm based on the filtered vibration and displacement data, wherein the compensation control parameters include adjustment values ​​of conveying speed and acceleration; The compensation control parameters are transmitted to the servo motor controller, and the vibration and displacement of the wafer conveying process are suppressed by adjusting the speed and acceleration of the servo motor; During the compensation control process, the vibration and displacement state of the wafer are continuously monitored, and the compensation control parameters are dynamically optimized through an adaptive control algorithm; When the wafer is stably transported to the rotary cleaning device, the compensation control is stopped and relevant data during the compensation control process is recorded.

6. The intelligent control method for a wafer cleaning and conveying process according to any one of claims 1 to 3, characterized in that: The S5 includes: A high-resolution camera is used to acquire images of the stationary wafer surface. The images are then converted to grayscale using an image preprocessing algorithm to enhance image contrast and provide high-quality input for subsequent image segmentation and feature extraction. The pre-processed image is segmented using an edge detection algorithm to accurately extract the wafer edge contour, and the wafer surface area is segmented using a region growing algorithm to obtain characteristic data of the initial state of the wafer surface. Extract feature information of surface defects and stains from the segmented image, compare it with the preset defect library using a feature matching algorithm, and determine the initial state classification label of the wafer surface; During the cleaning process, a high-speed camera is used to capture the instantaneous image of the rotating wafer in real time. The instantaneous image is aligned with the initial state image through an image registration algorithm to ensure the temporal and spatial consistency of the image analysis. Perform color space conversion on the registered instantaneous image to extract the color characteristics of the cleaning liquid on the wafer surface, and use the area coverage analysis algorithm to evaluate the coverage area and uniformity of the cleaning liquid; Based on the coverage area and uniformity analysis results, combined with the preset coverage scoring model, a quantitative coverage score is calculated. The scoring model takes into account the coverage area ratio and distribution uniformity indicators; If the coverage score is lower than the preset threshold, feedback is provided to adjust the cleaning liquid injection parameters and the cleaning process is repeated until the score reaches the preset standard to ensure that the wafer surface cleaning effect meets the process requirements.

7. The intelligent control method for a wafer cleaning and conveying process according to any one of claims 1 to 3, characterized in that: The S6 includes: Acquire real-time coverage score data, compare the real-time coverage score data with a preset threshold, and trigger a fuzzy control algorithm if the real-time coverage score data is lower than the preset threshold; Input the current coverage score, rotation angle and speed into the fuzzy control algorithm, calculate according to the preset fuzzy rules, and obtain the optimal angle adjustment amount and the optimal speed adjustment amount; Feeding back the optimal angle adjustment amount and the optimal speed adjustment amount to the cleaning equipment to dynamically adjust the wafer rotation parameters; Continuously monitoring coverage changes, and if the adjusted coverage score is still lower than the preset threshold, restarting the fuzzy control algorithm for optimization; Record the coverage score and corresponding rotation angle and speed parameters after each adjustment to construct an optimized data set; Iteratively improving the fuzzy rules in the fuzzy control algorithm according to the optimized data set to obtain an improved fuzzy control algorithm; The improved fuzzy control algorithm is used to optimize and control the subsequent wafer cleaning process, thereby improving the adaptability and accuracy of the cleaning effect.