Method for detecting and monitoring cleaning effect of semiconductor wafer in real time

Abstract

The invention relates to the technical field of information, in particular to a method for detecting and monitoring the cleaning effect of a semiconductor wafer in real time in the technical field of semiconductor wafer manufacturing, which comprises the following steps: introducing a rapid chemical analysis technology, combining an electrochemical sensor and spectral analysis, monitoring metal ion residues on the surface of the wafer in real time, and shortening the detection period; instant feedback on a production line is realized; a real-time feedback control system is constructed, according to the output of the comprehensive evaluation model, the working parameters of the cleaning equipment are automatically adjusted, the cleaning process is optimized, and the usage amount of cleaning liquid and the environmental influence are reduced; an edge computing technology is integrated, data processing and analysis tasks are sunk to a cleaning equipment end, data transmission delay is reduced, and the response speed of real-time monitoring and feedback control is increased; a visual monitoring platform is developed, key parameters, detection results and prediction trends in the cleaning process are displayed in a graphical mode, and operators can conveniently master the cleaning state in real time and quickly respond to abnormal conditions.

Description

Technical Field

[0001] The present invention relates to the field of information technology, specifically to the field of semiconductor wafer manufacturing technology, and in particular to a real-time detection and monitoring method for semiconductor wafer cleaning effects. Background Art

[0002] Cleaning is a key step in the semiconductor wafer manufacturing process, which directly affects the yield and performance of the final chip. The wafer surface needs to be inspected with high precision before and after cleaning to evaluate the cleaning effect and confirm whether the contaminants such as particles and metal ion residues on the wafer surface meet the process requirements.

[0003] Traditional optical detection methods have insufficient accuracy when dealing with sub-nanometer defect detection. Specifically, with increasingly stringent requirements for wafer surface cleanliness, existing detection methods are difficult to accurately quantify the number and size distribution of ultrafine particles, especially for particles smaller than 10 nanometers, whose scattering signals are extremely weak and easily overwhelmed by system noise. At the same time, for metal ion contamination, traditional methods rely on chemical analysis, which has a long detection cycle and cannot achieve real-time monitoring on the production line.

[0004] In addition, during the wafer cleaning process, slight fluctuations in parameters such as the composition, concentration, temperature, and cleaning time of the cleaning solution will affect the final cleaning effect, and thus affect the accuracy of the test results. However, existing detection devices are often not sensitive enough to monitor these process parameters, and cannot promptly feedback potential problems in the cleaning process, resulting in a reduction in detection accuracy, and cannot meet the current high-precision and high-efficiency semiconductor wafer cleaning test needs. Summary of the invention

[0005] The present invention provides a real-time detection and monitoring method for semiconductor wafer cleaning effect, the method comprising the following steps:

[0006] Step S101, using high-sensitivity nano-scale particle detection technology and optical sensing equipment to capture the scattered signals of sub-nanometer particles on the wafer surface, and improving the sensitivity and accuracy of particle detection through signal amplification and noise suppression algorithms;

[0007] Step S102: Introduce rapid chemical analysis technology, combine electrochemical sensors and spectral analysis, monitor the metal ion residues on the wafer surface in real time, shorten the detection cycle, and achieve instant feedback on the production line;

[0008] Step S103: Design an intelligent parameter monitoring system to collect parameters of the composition, concentration, temperature and cleaning time of the cleaning solution in real time, use a machine learning algorithm to analyze the impact of parameter fluctuations on the cleaning effect, and dynamically adjust the cleaning process parameters to ensure a stable cleaning effect;

[0009] Step S104: Develop a multimodal data fusion algorithm to integrate optical detection data, chemical analysis data, and cleaning parameter data, build a comprehensive evaluation model, comprehensively evaluate the wafer cleaning effect, and improve the reliability of the detection results;

[0010] Step S105: applying deep learning image processing technology to denoise, enhance and extract features from the captured wafer surface image, accurately identify and quantify nano-scale particles, and generate a high-resolution surface cleanliness distribution map;

[0011] Step S106: construct a real-time feedback control system to automatically adjust the working parameters of the cleaning equipment according to the output of the comprehensive evaluation model, optimize the cleaning process, and reduce the amount of cleaning fluid used and the environmental impact;

[0012] Step S107: design a data-driven cleaning effect prediction model, use historical cleaning data and real-time monitoring data to predict future cleaning effect trends, warn of potential problems in advance, and improve the preventive maintenance capabilities of the production line;

[0013] Step S108: Integrate edge computing technology to transfer data processing and analysis tasks to the cleaning equipment end, reduce data transmission delays, and improve the response speed of real-time monitoring and feedback control;

[0014] Step S109: Develop a visual monitoring platform to display key parameters, test results and predicted trends of the cleaning process in a graphical manner, so that operators can understand the cleaning status in real time and respond quickly to abnormal situations.

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

[0016] The present invention discloses a real-time detection and monitoring method for the cleaning effect of semiconductor wafers. The core business scenario problem is how to quickly and accurately analyze metal ion residues while detecting nano-scale particles with high sensitivity, and combine cleaning process parameters to achieve comprehensive evaluation and immediate optimization of the cleaning effect. In response to this problem, the present invention first uses high-sensitivity nano-scale particle detection technology and rapid chemical analysis technology to capture the particle and metal ion information on the wafer surface respectively; then, an intelligent parameter monitoring system is designed to collect cleaning process parameters in real time, and the detection data and process parameters are integrated through a multimodal data fusion algorithm; finally, a real-time feedback control system is constructed to automatically adjust the cleaning equipment parameters according to the results of the comprehensive evaluation model. At the same time, deep learning image processing technology is applied to accurately identify and quantify nano-scale particles, and a data-driven cleaning effect prediction model is used to predict future trends, and the response speed is improved by combining edge computing technology. Finally, the cleaning status is intuitively displayed through a visual monitoring platform. This solution significantly improves the controllability and reliability of the wafer cleaning process, optimizes the cleaning effect, and reduces resource consumption and environmental impact. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail with reference to the following specific embodiments.

[0018] The real-time detection and monitoring method of the semiconductor wafer cleaning effect of this embodiment may specifically include:

[0019] Step S101, using high-sensitivity nano-scale particle detection technology, using optical sensing equipment to capture the scattered signals of sub-nanoscale particles on the wafer surface, and improving the sensitivity and accuracy of particle detection through signal amplification and noise suppression algorithms.

[0020] Step S101 includes: obtaining a scattering signal from the surface of the wafer; if the light intensity distribution of the scattering signal conforms to the Gaussian distribution, amplifying the scattering signal to obtain an amplified scattering signal; if the light intensity distribution of the scattering signal does not conform to the Gaussian distribution, not amplifying the scattering signal; performing Fourier transform on the amplified scattering signal to obtain frequency domain data; if the proportion of high-frequency components in the frequency domain data exceeds a preset threshold, performing Gaussian filtering on the frequency domain data to obtain denoised frequency domain data; if the proportion of high-frequency components in the frequency domain data does not exceed the preset threshold, not performing Gaussian filtering on the frequency domain data; performing inverse Fourier transform on the denoised frequency domain data to obtain a time domain signal; determining the peak intensity and peak number of the scattering signal based on the peak characteristics of the time domain signal; determining the size and number of sub-nano particles on the surface of the wafer based on the peak intensity and peak number of the scattering signal.

[0021] Specifically, an optical microscope with a numerical aperture of 9 is used to capture the scattered signals generated by sub-nano particles on the surface of the wafer under irradiation with a laser of 405 nanometers. Assuming that the intensity distribution of the scattered light signal sampled at a certain time is {10, 12, 15, 18, 20, 18, 15, 12, 10}, this set of data approximately conforms to the Gaussian distribution. The signal amplifier amplifies the signal with an amplification factor of 100 times, and the intensity distribution of the amplified signal is {1000, 1200, 1500, 1800, 2000, 1800, 1500, 1200, 1000}. The amplified scattered signal is subjected to a fast Fourier transform to obtain frequency domain data. Assuming that in the frequency domain data, the total energy is 10,000, and the energy of the high-frequency component above 10 kHz is 6,000, accounting for 60%, which exceeds the preset threshold of 50%. Therefore, the Gaussian filtering algorithm is used for noise suppression. The standard deviation of the Gaussian filter is set to 2 kHz, and the frequency domain data is filtered to obtain the denoised frequency domain data. The denoised frequency domain data is inversely Fourier transformed to obtain the time domain signal. The time domain signal is processed by the peak finding algorithm, the threshold of the peak finding algorithm is set to 1500, the spacing is set to 3, and the peak intensity of the scattering signal is obtained as {1800, 2000, 1800}, and the number of peaks is 3. Assuming that the peak intensity is proportional to the cube of the particle size, the number of peaks is proportional to the number of particles, and it is known that the peak intensity of 2000 corresponds to 50 nanometer particles, the particle size corresponding to other peak intensities can be inferred, for example, the peak intensity of 1800 corresponds to approximately 45 nanometer particles. Based on the number of peaks of 3, it can be inferred that there are 3 sub-nanometer particles on the wafer surface.

[0022] Step S102, introduce rapid chemical analysis technology, combine electrochemical sensors and spectral analysis, monitor the metal ion residues on the wafer surface in real time, shorten the detection cycle, and achieve instant feedback on the production line.

[0023] Step S102 includes: obtaining a set of original electrical signals of the surface layer of the wafer, performing Fourier transform on the original electrical signal set to obtain electrical signal frequency domain data; obtaining spectral data of the surface layer of the wafer, and obtaining spectral absorbance by least squares calculation based on the spectral data; determining the type of residue information by comparing the spectral absorbance with a pre-established metal ion bond absorbance database; obtaining the residue type information and the electrical signal frequency domain data, and obtaining fused data based on the residue type information and the electrical signal frequency domain data; inputting the fused data into a trained artificial neural network model to obtain a predicted value of metal ion concentration in the surface layer; comparing the predicted value of metal ion concentration in the surface layer with a preset threshold value, and generating an alarm signal if the predicted value is greater than the threshold value; obtaining the alarm signal, and controlling the production line to suspend operation based on the alarm signal; generating a visualization report based on the alarm signal, and the report includes residue type, residue concentration, and residue location information.

[0024] Specifically, an array of 32 electrochemical sensors is used to perform grid scanning on the surface of the wafer. Each sensor stays at a specific position for 5 seconds to collect electrical signals, and finally obtains 32 sets of original electrical signal sets, recorded as {S1, S2, ..., S32}. A fast Fourier transform (FFT) is performed on each set of original electrical signals Si, and the sampling frequency is set to 1000 Hz to obtain the corresponding frequency domain data F i. For example, F1 contains a series of amplitude information at frequencies such as {f1 (10 Hz) = 1 volt, f1 (50 Hz) = 0.5 volt, ..., f1 (500 Hz) = 0.1 volt}, representing the intensity of different frequency components. At the same time, a spectrometer with a wavelength range of 200 nanometers to 800 nanometers is used to scan the surface of the wafer point by point to obtain spectral data at each position. For the spectral data at each position, the least square method is used for curve fitting, and the fitting function is a Gaussian function, and the formula is: A(λ)=a*exp(-(λ-b)^2 / (2*c^2)), wherein A(λ) is the absorptivity at wavelength λ, and a, b, and c are coefficients to be determined. The spectral absorptivity curve of each position is calculated by the least square method. The calculated spectral absorptivity curve is compared with a pre-established metal ion bond absorptivity database, which contains information such as copper ions having an absorption peak at 280 nanometers and iron ions having an absorption peak at 340 nanometers. The correlation coefficient is used as a similarity index to calculate the correlation coefficient of the spectrum to be measured and each spectrum in the database. For example, the correlation coefficient of the spectrum to be measured with the copper ion spectrum is 95, and the correlation coefficient with the iron ion spectrum is 6, then the residue type is determined to be copper ion. The residue type information (for example, copper ion is encoded as 001) is fused with 32 groups of frequency domain data F i to form a 35-dimensional feature vector as the input of the artificial neural network model. The artificial neural network model adopts a three-layer fully connected structure, with 35 nodes in the input layer, 64 nodes in the hidden layer, and 1 node in the output layer. The activation function uses the rectified linear unit (ReLU) and the loss function uses the mean square error (MSE). After training with a large amount of sample data, the model can predict the concentration value of copper ions in the surface layer based on the input feature vector, for example, the predicted value is 150 nanograms per square centimeter. The threshold is set to 100 nanograms per square centimeter. Since the predicted value of 150 nanograms per square centimeter is greater than the threshold, the system triggers an alarm signal. The alarm signal is transmitted to the control center via a wireless network. After receiving the alarm signal, the control center immediately sends a stop command to the production line to suspend the production line. At the same time, the control center generates a visual report based on the alarm signal. The report shows the distribution of copper ion concentration on the surface of the wafer in the form of a heat map. For example, at the coordinate (10, 20), the copper ion concentration is 150 nanograms per square centimeter and is marked in red. At the same time, the report indicates that the type of residue is copper ion, the concentration is 150 nanograms per square centimeter, and the position is (10, 20).Because the alarm signal contains location information, the control center can synchronously generate a production process inspection report, which will call the production record database to query the production step corresponding to the coordinates (10, 20). For example, the step is the copper interconnection process, the time is 10:30 on October 26, 2023, the operator is Zhang San, and the equipment number is EQP001. It will also analyze whether there is any abnormality with the equipment or the operator based on historical data, providing a basis for subsequent process improvements.

[0025] Step S103, design an intelligent parameter monitoring system to collect the parameters of the composition, concentration, temperature and cleaning time of the cleaning liquid in real time, use machine learning algorithms to analyze the impact of parameter fluctuations on the cleaning effect, dynamically adjust the cleaning process parameters, and ensure a stable cleaning effect.

[0026] Acquire cleaning liquid component data, the cleaning liquid component data including first component information, second component information, and third component information; determine a first concentration value of the first component information, a second concentration value of the second component information, and a third concentration value of the third component information according to the cleaning liquid component data; acquire a cleaning liquid temperature value, a cleaning time value, and an effect value obtained by comparing the surface cleanliness of the workpiece before and after cleaning; the first concentration value, the second concentration value, the third concentration value, the cleaning liquid temperature value, the cleaning time value, and the effect value constitute a first data set; use a support vector machine classifier to train the first data set, and obtain multiple range intervals of the first concentration value, the second concentration value, the third concentration value, the cleaning liquid temperature value, and the cleaning time value according to different effect values; if the first concentration value, the second concentration value, the third concentration value, the cleaning liquid temperature value, and the cleaning time value collected in real time exceed the range interval, obtain the first concentration value fluctuation rate, the second concentration value fluctuation rate, the third concentration value fluctuation rate, the cleaning liquid temperature value fluctuation rate, and the cleaning time value fluctuation rate according to the difference between the first concentration value, the second concentration value, the third concentration value, the cleaning liquid temperature value, and the cleaning time value and the range interval; the first concentration value The second data set is composed of a first concentration value volatility, a second concentration value volatility, a third concentration value volatility, a cleaning liquid temperature value volatility, a cleaning time value volatility, and an effect value; the second data set is trained with a random forest regressor to obtain a functional relationship between volatility and effect value, Y=F(x1, x2, x3, x4, x5), where Y is the effect value, x1 is the first concentration value volatility, x2 is the second concentration value volatility, x3 is the third concentration value volatility, x4 is the temperature value volatility, and x5 is the time value volatility; if the effect value is lower than a preset threshold, query the process table to determine the first concentration value volatility. A concentration value adjustment amount, a second concentration value adjustment amount, a third concentration value adjustment amount, a cleaning liquid temperature value adjustment amount, and a cleaning time value adjustment amount; according to the first concentration value adjustment amount, the second concentration value adjustment amount, the third concentration value adjustment amount, the cleaning liquid temperature value adjustment amount, and the cleaning time value adjustment amount, an adjusted process parameter is obtained; according to the adjusted process parameter, the actuator is controlled to adjust the parameters, and the first concentration value, the second concentration value, the third concentration value, the cleaning liquid temperature value, and the cleaning time value are changed to obtain a new effect value; and it is determined whether the new effect value is higher than a preset threshold.

[0027] Specifically, it is assumed that the cleaning liquid component table contains three main components: sodium hydroxide, sodium carbonate, and trisodium phosphate, which correspond to the first component, the second component, and the third component, respectively. Deploy multiple sensors, including a sodium hydroxide concentration sensor, a sodium carbonate concentration sensor, a trisodium phosphate concentration sensor, a temperature sensor, and a timer. For example, the initial setting of sodium hydroxide concentration is 5%, sodium carbonate concentration is 3%, trisodium phosphate concentration is 2%, the cleaning liquid temperature is 60 degrees Celsius, and the cleaning time is 10 minutes. The cleanliness meter detects that the surface cleanliness of the workpiece before cleaning is 10GU, and the cleanliness after cleaning is 95GU, then the effect value is 85GU. These data (sodium hydroxide concentration 5%, sodium carbonate concentration 3%, trisodium phosphate concentration 2%, temperature 60 degrees Celsius, time 10 minutes, effect value 85GU) are taken as part of the first data set. After collecting a large amount of similar data, a support vector machine (SVM) classifier is used for training. For example, data with an effect value greater than 80GU are marked as "excellent", 60-80GU are marked as "good", and data below 60GU are marked as "unqualified". The SVM classifier separates samples of different categories by constructing a hyperplane and finds the optimal classification boundary. After training, the range of parameters corresponding to different effect values ​​is obtained. For example, the sodium hydroxide concentration range corresponding to "excellent" is 5%-5%, the sodium carbonate concentration range is 8%-2%, the trisodium phosphate concentration range is 8%-2%, the temperature range is 55-65 degrees Celsius, and the time range is 9-11 minutes. When the real-time monitored sodium hydroxide concentration becomes 6%, the sodium carbonate concentration becomes 5%, the trisodium phosphate concentration becomes 3%, the temperature becomes 68 degrees Celsius, and the time becomes 8 minutes, it exceeds the above range. Calculate the volatility of each parameter: the volatility of sodium hydroxide concentration is (6%-5%) / 5%=20%, the volatility of sodium carbonate concentration is (5%-3%) / 3%=17%, the volatility of trisodium phosphate concentration is (3%-2%) / 2%=15%, the temperature volatility is (68-60) / 60=13%, and the time volatility is (10-8) / 10=20%. Assume that the effect value is 70GU at this time. These volatility data (20%, 17%, 15%, 13%, 20%, 70GU) are used as part of the second data set. After collecting a large amount of similar data, a random forest regressor is used for training. By constructing multiple decision trees and combining their prediction results, the functional relationship between volatility and effect value Y=F(x1,x2,x3,x4,x5) is obtained. Assuming that the effect value is lower than the preset threshold of 80GU, the pre-established process table is queried. The table is based on historical data and expert experience and records the adjustment amount of each parameter under different circumstances. For example, when the effect value is 70GU, the process table recommends increasing the concentration of sodium hydroxide by 5%, the concentration of sodium carbonate by 3%, the concentration of trisodium phosphate by 2%, the temperature by 3 degrees Celsius, and the time by 1 minute.According to the recommendations of the process table, the adjusted process parameters are: sodium hydroxide concentration 5%, sodium carbonate concentration 3%, trisodium phosphate concentration 2%, temperature 57 degrees Celsius, time 11 minutes. By controlling the actuators, such as the automatic feeding system and the heating system, the cleaning liquid is adjusted to reach the new process parameters. Clean again and test the effect value. Assuming that the new effect value is 83GU, which is higher than the preset threshold of 80GU, the adjustment is considered successful. If the effect value is still lower than the preset threshold, it is necessary to continue to adjust the parameters or manually intervene to analyze the cause.

[0028] Step S104, develop a multimodal data fusion algorithm to integrate optical inspection data, chemical analysis data and cleaning parameter data, build a comprehensive evaluation model, comprehensively evaluate the wafer cleaning effect, and improve the reliability of the inspection results.

[0029] The image information output by the optical detection device is obtained, and the scattering state information of the wafer surface is extracted by using an image processing algorithm to obtain a first scattering feature vector; the first scattering feature vector and the metal residue on the wafer surface output by the chemical analyzer are obtained, and a data normalization algorithm is used for normalization processing to obtain the first fused data; according to the first fused data and the cleaning pressure value output by the cleaning device, a principal component analysis algorithm is used for dimension reduction to obtain the second fused data; according to the second fused data and the cleaning temperature value output by the cleaning device, a random forest regression model is constructed to obtain a wafer cleanliness prediction value; the wafer cleanliness prediction value and the cleaning solvent type output by the cleaning device are obtained, and a support vector machine classification model is constructed to obtain a wafer cleaning effect level; according to the wafer cleaning effect level and the cleaning time output by the cleaning device, an XGBoost regression model is constructed to obtain an optimized cleaning time; according to the optimized cleaning time and the number of particles on the wafer surface output by the optical detection device, a closed-loop feedback control system is constructed. If the number of particles on the wafer surface is greater than a preset threshold, the cleaning time is adjusted; if the number of particles on the wafer surface is not greater than the preset threshold, the cleaning process is determined to be ended.

[0030] Specifically, the optical detection equipment scans the wafer surface to generate a grayscale image with a resolution of 1024x1024 pixels. The Laplace operator is used for edge detection to extract features such as the number of scattering points, average scattering intensity, and scattering distribution uniformity to form a first scattering feature vector containing 15 eigenvalues ​​[2,5,…,8]. The first scattering feature vector is merged with the data of the three metal residues of copper, iron, and nickel on the wafer surface detected by the chemical analyzer [1ppm, 05ppm, 02ppm]. The Z-score standardization method is used, that is, each data point is subtracted from the mean and then divided by the standard deviation to obtain the first fused data [-2,5,…,8,-8,3,1]. The first fused data is merged with the cleaning pressure value of 50psi recorded by the cleaning equipment. The principal component analysis algorithm is used to set the principal component contribution rate threshold to 95%, and the data is reduced to 5 dimensions to obtain the second fused data [1,-8,…,5]. The second fusion data is merged with the cleaning temperature value of 60 degrees Celsius recorded by the cleaning equipment. A random forest regression model containing 100 decision trees is used to input the historical data training model, and the wafer cleanliness value is predicted to be 95%. The predicted value of wafer cleanliness 95% is merged with the cleaning solvent type diisopropyl alcohol used by the cleaning equipment. The support vector machine classification model is used to divide the wafer cleaning effect level into four levels: excellent, good, medium, and poor, and the current cleaning effect level is determined to be "excellent". Using the XGBoost regression model, according to the cleaning time data corresponding to the "excellent" cleaning effect level in the historical data, the optimized cleaning time is predicted to be 120 seconds. A closed-loop feedback control system is constructed, and the optimized cleaning time of 120 seconds is set as the cleaning time parameter of the cleaning equipment. After the cleaning is completed, the optical detection equipment detects the number of particles on the wafer surface again as 30, and the preset threshold is 20. It is determined that it is greater than the preset threshold. The system automatically increases the cleaning time by 10 seconds, cleans again, and cyclically detects and adjusts until the number of particles on the wafer surface is detected to be reduced to 18, which is less than the preset threshold of 20, and the cleaning process is determined to be over.

[0031] Step S105, applying deep learning image processing technology to denoise, enhance and extract features from the captured wafer surface image, accurately identify and quantify nano-scale particles, and generate a high-resolution surface cleanliness distribution map.

[0032] The invention provides a method for collecting an original image of a wafer surface; performing a convolution operation on the original image using a pre-established convolutional neural network model to obtain a first image; setting Gaussian filter parameters according to the noise level of the first image; performing a Gaussian filter operation on the first image using the Gaussian filter parameters to obtain a denoised second image; performing grayscale mapping on the second image using a histogram equalization algorithm according to the contrast of the second image to determine an enhanced third image; obtaining gradient information of the third image; performing edge detection processing on the third image using a preset Canny edge detection operator to obtain a fourth image, wherein the fourth image contains edge information; selecting a morphological processing method according to the edge information of the fourth image to perform particle segmentation to obtain a segmented fifth image, wherein the fifth image contains particle contour information; obtaining particle size distribution data by calculating the pixel area of ​​each particle contour area for the particle contour information contained in the fifth image; calculating the number of pixels of each particle to obtain particle quantity data; mapping the particle size distribution data and the particle quantity data to corresponding spatial coordinates to generate a surface cleanliness distribution map.

[0033] Specifically, suppose we have collected a 512x512 pixel original grayscale image of the wafer surface, with a pixel value range of 0-255. First, we can use a pre-trained convolutional neural network model for wafer surface defect detection tasks, such as a network containing 5 convolutional layers and 2 fully connected layers, each convolutional layer is followed by a ReLU activation function and a maximum pooling layer. Input the original image into the model, and after forward propagation, obtain the first image, assuming that the size of the first image is 128x128. Then, analyze the noise level of the first image, for example, calculate the signal-to-noise ratio of the image, assuming that the signal-to-noise ratio is 20 decibels. Based on this signal-to-noise ratio, we can set a 5x5 Gaussian filter with a standard deviation of 5, and perform a convolution operation on the first image to filter out high-frequency noise to obtain the denoised second image. Next, calculate the grayscale histogram of the second image, and find that its grayscale levels are mainly concentrated between 50-150, with low contrast. In order to enhance the contrast, we can use the histogram equalization algorithm, calculate the cumulative distribution function, and map the original grayscale to a new grayscale range of 0-255 to obtain the enhanced third image. Then, calculate the gradient of the third image in the horizontal and vertical directions, for example, use the Sobel operator to obtain the gradient amplitude image. Based on the gradient amplitude image, we use the preset Canny edge detection operator, where the high threshold is set to 80 and the low threshold is set to 30, to perform edge detection and obtain the fourth image containing edge information. Observing the fourth image, it is found that there are breaks in the edges of the particles, so the morphological closing operation is selected, using a circular structure element with a radius of 2, first performing an expansion operation, and then performing an erosion operation to connect the broken edges to obtain the segmented fifth image. Each closed area in the fifth image represents a particle. Finally, the fifth image is marked with connected areas, and the number of pixels in each connected area, that is, the pixel area, is calculated. Assuming that 100 particles are obtained, the areas are 10, 25, 15, etc. pixel units, corresponding to the particle size. At the same time, the number of pixels in each particle area is counted to obtain the particle number data. The size and quantity information of these particles are mapped to a new 512x512 blank image according to their coordinate positions in the original image. For example, areas with particle sizes greater than 20 pixels are marked in red, and areas with a quantity greater than 5 are marked in blue, ultimately generating a high-resolution surface cleanliness distribution map.

[0034] Step S106, constructing a real-time feedback control system, automatically adjusting the working parameters of the cleaning equipment according to the output of the comprehensive evaluation model, optimizing the cleaning process, and reducing the amount of cleaning fluid used and the environmental impact.

[0035] Step S106 also includes the following steps:

[0036] Step 1: Collect the surface image information of the cleaning object, extract the image feature information through the image recognition algorithm, and obtain the initial value of the cleanliness of the cleaning object. Step 2: According to the initial value of the cleanliness of the cleaning object, match the pre-established database of cleaning liquid concentration, temperature and cleaning time to determine the initial working parameter combination of the cleaning equipment. Step 3: The cleaning equipment operates under the initial working parameter combination, and the sensor monitors the turbidity and conductivity index data of the discharged waste liquid during the cleaning process to obtain real-time data. Step 4: Compare the real-time data with the cleaning liquid concentration, temperature and cleaning time data under the initial value of the cleanliness of the cleaning object corresponding to the historical data. If the deviation between the real-time data and the historical data exceeds the preset threshold, the adjustment strategy is triggered. Step 5: The adjustment strategy uses the long short-term memory network algorithm to predict the change trend of the cleanliness of the cleaning object according to the direction and amplitude of the deviation between the real-time data and the historical data, and obtains the predicted final cleanliness value. Step 6: Determine whether the predicted final cleanliness value matches the preset cleanliness target value. If not, use the gradient boosting decision tree algorithm to output the equipment parameter adjustment instruction to obtain the adjustment amount of the equipment parameter. Step 7: The control system receives the adjustment amount of the equipment parameters, adjusts the cleaning liquid concentration, temperature and cleaning time parameters of the cleaning equipment, and executes step 3 repeatedly until the predicted final cleanliness value matches the preset cleanliness target value to obtain an optimized cleaning process.

[0037] Specifically, assume that the surface image information of an electronic component is collected, and the image resolution is 1280x720 pixels. The feature extraction is performed through a pre-trained convolutional neural network (such as ResNet-50), and a 2048-dimensional feature vector is output. Then, the feature vector is mapped to a cleanliness level using a support vector machine classifier. For example, the cleanliness is divided into 1-10 levels, and the calculated initial cleanliness value of the electronic component is level 3, which is a relatively dirty state. According to the initial cleanliness value of 3, it is matched in a pre-established database, and the database stores the empirical values ​​of the cleaning solution concentration, temperature and cleaning time corresponding to different cleanliness levels. For example, the cleaning solution concentration corresponding to cleanliness level 3 is 5%, the temperature is 60 degrees Celsius, and the cleaning time is 15 minutes. These values ​​are used as the initial working parameter combination of the cleaning equipment. The cleaning equipment operates according to the initial parameters, and the cleaning solution concentration is 5% sodium hydroxide aqueous solution, the temperature is 60 degrees Celsius, and the cleaning time is 15 minutes. During the cleaning process, the turbidity sensor and the conductivity sensor monitor the turbidity and conductivity of the discharged waste liquid in real time, and collect data once per second. For example, at the 5th minute, the turbidity sensor reading is 50NTU, and the conductivity sensor reading is 1500μS / cm. These real-time data are compared with historical data. In the historical data, the electronic components with an initial cleanliness value of 3 have a turbidity mean of 45NTU and a standard deviation of 5NTU, and a conductivity mean of 1400μS / cm and a standard deviation of 100μS / cm at the 5th minute under the same cleaning parameters. Although the current real-time turbidity data of 50NTU is within the range of three times the standard deviation of the historical data, the conductivity data of 1500μS / cm exceeds the range of the mean of the historical data plus three times the standard deviation (1400+3*100=1700μS / cm), that is, the deviation exceeds the preset threshold, triggering the adjustment strategy. According to the deviation between real-time data and historical data, the long short-term memory network (LSTM) algorithm is used to predict the cleanliness change trend. The LSTM model takes the cleanliness, turbidity, and conductivity of the previous moment as input and outputs the cleanliness prediction value of the next moment. For example, the cleanliness prediction value at the current moment is 4, the turbidity is 50NTU, and the conductivity is 1500μS / cm. The input is the LSTM model, and the cleanliness value at the next moment is predicted to be 2. Repeat this process until the cleaning time is over, and the predicted cleanliness final value is 6. Assuming that the preset cleanliness target value is 8, the predicted cleanliness final value 6 does not match it, and the gradient boosting decision tree (GBDT) algorithm is used to output the equipment parameter adjustment instructions. The GBDT model takes the current cleanliness, turbidity, conductivity, cleaning fluid concentration, temperature, and cleaning time as input, and outputs the adjustment amount of cleaning fluid concentration, temperature, and cleaning time.For example, the current cleanliness is 6, the turbidity is 30NTU, the conductivity is 1200μS / cm, the cleaning fluid concentration is 5%, the temperature is 60 degrees Celsius, and the cleaning time is 5 minutes. The GBDT model is input and the output cleaning fluid concentration adjustment is +1%, the temperature adjustment is +5 degrees Celsius, and the cleaning time adjustment is +3 minutes. The control system receives the adjustment instruction, adjusts the cleaning fluid concentration to 6%, the temperature to 65 degrees Celsius, and extends the cleaning time by 3 minutes. Continue to execute the cleaning process and repeat the above monitoring, comparison, prediction and adjustment steps until the final cleanliness value predicted by the LSTM model reaches or exceeds 8, for example, reaches 5. At this time, it is considered that the cleaning process has reached the optimized state and the cleaning is completed.

[0038] Step S107, designing a data-driven cleaning effect prediction model, using historical cleaning data and real-time monitoring data to predict future cleaning effect trends, early warning of potential problems, and improving the preventive maintenance capabilities of the production line.

[0039] Acquire a historical data set and a real-time data set, wherein the historical data set includes production data, equipment status, process parameters, environmental parameters, quality indicators, and historical cleaning records, and the real-time data set includes production data, equipment status, process parameters, environmental parameters, quality indicators, and real-time monitoring information; perform time alignment according to the time series information of the historical data set and the real-time data set to obtain a total data set; remove missing values ​​and outliers in the total data set to obtain a complete data set; perform data dimension reduction using principal component analysis according to the production data, equipment status, process parameters, and environmental parameters in the complete data set to obtain a reduced-dimensional data set; divide the reduced-dimensional data set into a training set and a test set, wherein the training set includes historical cleaning records, and the test set includes real-time monitoring information; construct a time series prediction model based on a long short-term memory network according to the training set, wherein the prediction model includes an input layer, a hidden layer, and an output layer, wherein the input layer receives the training set, the hidden layer includes multiple neurons, and the output layer outputs a predicted cleaning effect; calculate the root mean square error RMSE between the predicted cleaning effect output by the prediction model and the actual cleaning effect according to the test set, and the formula is Among them, n is the number of test set samples, y_i is the actual cleaning effect of the i-th sample, is the predicted cleaning effect of the ith sample; if the root mean square error RMSE is greater than a preset threshold, the model parameters of the long short-term memory network are adjusted until the root mean square error RMSE is less than or equal to the preset threshold, and the trained long short-term memory network is determined to be the final prediction model; according to the time series information of the real-time monitoring information, the real-time monitoring information is time-sliced ​​to obtain a plurality of time slice data; through the time slice data, the final prediction model is used to obtain the predicted value of the cleaning effect corresponding to each time slice data in the future set time period; for the predicted value of the cleaning effect corresponding to each time slice data, the first-order difference and the second-order difference of the predicted value of the cleaning effect are calculated, the first-order difference represents the changing speed of the cleaning effect, and the second-order difference represents the changing acceleration of the cleaning effect; according to the first-order difference and the The second-order difference is used to judge the changing trend of the cleaning effect and obtain the cleaning effect trend information; if the cleaning effect prediction value is lower than the quality indicator threshold and the cleaning effect trend information shows a continuous downward trend, an early warning signal is generated, and the early warning signal includes an early warning level and recommended maintenance measures, and the early warning level is determined according to the degree to which the cleaning effect prediction value deviates from the quality indicator threshold, and the recommended maintenance measures are determined according to the cleaning effect trend information; according to the early warning level in the early warning signal, the priority of the maintenance task is determined, and the priority of the maintenance task includes three levels: high, medium and low; according to the recommended maintenance measures in the early warning signal, a maintenance work order is generated, and the maintenance work order includes a maintenance task description, maintenance time, maintenance personnel, and maintenance resources; through the maintenance work order, preventive maintenance of the production line equipment is carried out to obtain the equipment status data after maintenance.

[0040] Specifically, it is assumed that the time span of historical data is from January 1, 2022 to December 31, 2023, and the time span of real-time data is from January 1, 2024 to January 15, 2024. Both historical data and real-time data are recorded in hours. First, the historical data and real-time data are sorted by timestamp, and then the two data sets are merged based on the timestamp. For example, the historical data from 0:00 on January 1, 2022 to 23:00 on December 31, 2023 is merged with the real-time data from 0:00 on January 1, 2024 to 23:00 on January 15, 2024 to obtain the total data set. The total data set is preprocessed, and the missing values ​​in the total data set are filled by linear interpolation. For example, if the temperature data of a certain hour is missing, the average value of the temperature data of the two hours before and after the hour is used as the filling value. For abnormal values, the 3σ principle is used for identification and processing. For example, if the pressure value of a certain hour exceeds the range of the mean plus or minus 3 times the standard deviation, the pressure value is replaced by the boundary value of the mean plus or minus 3 times the standard deviation. After preprocessing, a complete data set is obtained. For the complete data set, production data (such as output, production speed), equipment status (such as temperature, pressure, vibration frequency), process parameters (such as feed rate, reaction time), and environmental parameters (such as ambient temperature and humidity) are selected as characteristic variables, and the principal component analysis method is used for dimensionality reduction. First, the correlation coefficient matrix between each characteristic variable is calculated, and then the eigenvalue and eigenvector of the correlation coefficient matrix are calculated. The first few principal components with a cumulative contribution rate of more than 85% are selected as new characteristic variables. For example, there are 10 original characteristic variables. After principal component analysis, the first 3 principal components are selected, and their cumulative contribution rate reaches 90%. Then the original 10 characteristic variables are reduced to 3 principal components. The reduced-dimensionality dataset is divided into a training set and a test set in a ratio of 8:2, where the training set contains data from January 1, 2022 to September 30, 2023, and the test set contains data from October 1, 2023 to December 31, 2023. The training set also includes historical cleaning records, such as the time of each cleaning, the type of cleaning agent, and the cleaning effect score. A long short-term memory network model is constructed, with 3 neurons in the input layer, corresponding to 3 principal components, two hidden layers, each containing 64 neurons, and 1 neuron in the output layer, which outputs the predicted cleaning effect score, ranging from 0 to 10 points. The long short-term memory network model is trained using the training set, and the performance of the model is evaluated using the test set. The root mean square error RMSE between the predicted cleaning effect score and the actual cleaning effect score is calculated. For example, there are 100 samples in the test set, and the root mean square error (RMSE) between the predicted value and the actual value of the cleaning effect score of these 100 samples is 15, which is greater than the preset threshold of 1. In this case, the parameters of the long short-term memory network model are adjusted, such as increasing the number of hidden layer neurons or adjusting the learning rate, and the model is retrained until the RMSE is less than or equal to 1.Assuming that the final RMSE is 0.8, the model is determined to be the final prediction model. Obtain real-time monitoring data from 0:00 on January 16, 2024 to 23:00 on January 16, 2024, and slice the data into 4 time slices every 6 hours. Input each time slice data into the trained prediction model to obtain the cleaning effect prediction value corresponding to each time slice, which is 5 points, 2 points, 8 points, and 5 points respectively. Calculate the first-order difference and second-order difference of the cleaning effect prediction value. The first-order difference is -3, -4, and -3 respectively, and the second-order difference is -1 and 1 respectively. According to the first-order difference and the second-order difference, it can be judged that the cleaning effect is on a downward trend, and the decline rate shows signs of accelerating. Set the quality index threshold to 7 points. If the cleaning effect prediction value is lower than 7 points and continues to decline, an early warning signal is generated. Generate an early warning signal based on the cleaning effect prediction value and trend information. For example, the cleaning effect prediction value of the third time slice is 8 points, which is lower than the threshold of 7 points, and the first-order difference is negative and the second-order difference is positive, indicating that the cleaning effect continues to decline, and an early warning signal is generated. The early warning level is determined according to the degree to which the cleaning effect prediction value deviates from the threshold. For example, the difference between 8 and 7 is 2 points, and the early warning level is medium. The recommended maintenance measures are determined based on the trend information. For example, if the decline rate is accelerating, it is recommended to perform cleaning maintenance immediately. Generate a maintenance work order, the maintenance task description is "production line equipment cleaning and maintenance", the maintenance time is "January 16, 2024 18:00", the maintenance personnel is "maintenance team A", and the maintenance resources are "cleaning agent X, 10 liters". After the maintenance is completed, the equipment status data after maintenance, such as temperature, pressure, vibration frequency, etc., are recorded, and the equipment status data after maintenance is added to the real-time data, and the model is updated to improve the prediction accuracy.

[0041] Step S108, integrating edge computing technology, sinking data processing and analysis tasks to the cleaning equipment end, reducing data transmission delays, and improving the response speed of real-time monitoring and feedback control.

[0042] The original data stream collected by the device is obtained to generate an initial data set, and the initial data set is transmitted to the edge side through wireless communication. The edge side receives the initial data set to obtain a first data set; if the size of the first data set exceeds a preset threshold value A, the first data set is compressed to obtain a second data set. If the size of the first data set does not exceed the preset threshold value A, the first data set is used as the second data set; according to the size of the second data set, a second data set transmission protocol is determined; according to the second data set transmission protocol, the second data set is transmitted to the edge server using the UDP protocol, and the edge server receives the second data set to obtain a third data set; data cleaning operations are performed on the third data set to remove the data that do not meet the requirements of the third data set. The fourth data set is obtained by combining the data that meets the preset rule B; the feature extraction algorithm is used to extract the features of the fourth data set to obtain a feature set; according to the preset rule B, the decision tree algorithm is used to classify the feature set, and the category to which each feature belongs is determined to obtain a classified feature set; through the classified feature set, the K-nearest neighbor algorithm is used to perform anomaly detection to determine whether there is abnormal data; if there is abnormal data, an abnormal alarm information is generated; if there is no abnormal data, the classified feature set is used as a normal data set; according to the normal data set and the abnormal alarm information, a prediction model is constructed by using a logistic regression algorithm, and the prediction model outputs the equipment operation status, and according to the equipment operation status, it is determined whether to remotely control the cleaning equipment.

[0043] Specifically, multiple sensors, such as pressure sensors and temperature sensors, are installed on the intelligent cleaning equipment. Data is collected 10 times per second, and the size of each collected data is 1KB. The data is collected continuously for 1 minute to generate a raw data stream of 600KB. This 600KB raw data stream is packaged into an initial data set, and the initial data set is transmitted to the edge side via 4G wireless communication. After the edge side receives the initial data set, it adds 5KB of transmission information to obtain a first data set of 605KB. The preset threshold A is 500KB. Since the size of the first data set of 605KB is greater than the preset threshold of 500KB, the LZ77 algorithm is used to compress the first data set with a compression rate of 50%, and a second data set of 305KB is obtained. The transmission protocol of the second data set is determined to be UDP, and 5KB of protocol information is added, and the data packet size becomes 305KB. The edge server receives the second data set transmitted via the UDP protocol, removes 5KB of protocol information, and obtains a third data set of 305KB. The preset rule B is: the pressure value range is between 2MPa and 8MPa, and the temperature value range is between 20 degrees Celsius and 80 degrees Celsius. Traverse each data point in the third data set to check whether its pressure value and temperature value meet the preset rule B. For example, if the pressure value of a data point is 1MPa and does not meet the rule, it will be deleted from the third data set. After data cleaning, a fourth data set with a size of 280KB is obtained. The principal component analysis algorithm is used to extract features from the fourth data set, and the original 10 features (such as pressure, temperature, flow, etc.) are reduced to 5 principal components to obtain a feature set containing 5 features. According to the preset rule B, that is, the range of pressure values ​​and temperature values, the classification rules of the decision tree algorithm are set. For example, when the pressure value is greater than 5MPa and the temperature value is greater than 50 degrees Celsius, the feature is classified into the "high pressure and high temperature" category. Through the decision tree algorithm, each feature in the feature set is classified to obtain a classified feature set. In the classified feature set, the K-nearest neighbor algorithm is used for anomaly detection, and the K value is set to 5, that is, to find the 5 nearest neighbors of each data point. If the average distance between a data point and its five neighbors is greater than a preset threshold (for example, 5 times the standard deviation), the data point is judged to be abnormal data, and an abnormal alarm message is generated, such as "device A has abnormal pressure at time point T". If the average distance between a data point and its five neighbors is less than or equal to the preset threshold, the data point and its category are added to the normal data set. Assume that the normal data set contains 1,000 data points and the abnormal alarm message contains 10 records. Using the logistic regression algorithm, the normal data set and the abnormal alarm message are used as training data to construct a prediction model. Among them, the label of the normal data set is 0, and the label of the abnormal alarm message is 1. Through training, the parameters of the logistic regression model are obtained. When a new data point is input, the prediction model outputs the probability that the data point belongs to an abnormal state.For example, if the output probability is 8, which exceeds the preset threshold of 6, the predicted equipment operation status is abnormal, and remote control of the cleaning equipment can be triggered, such as sending instructions to reduce the equipment operating power or stop the equipment. If no abnormality is triggered, an instruction is sent to maintain the current equipment operation status.

[0044] Step S109, develop a visual monitoring platform to display the key parameters, test results and predicted trends of the cleaning process in a graphical manner, so that operators can understand the cleaning status in real time and respond quickly to abnormal situations.

[0045] Acquire equipment operation data, the equipment operation data including vibration data, temperature data and pressure data collected by sensors; perform data format conversion based on the equipment operation data to obtain a pre-processed data set in a unified data format; calculate statistical indicators of key parameters based on the pre-processed data set, the statistical indicators including mean and standard deviation; apply a pre-established random forest algorithm model based on the statistical indicators of the key parameters to perform cleaning state identification, and obtain a cleaning state label corresponding to each data in the pre-processed data set; construct a prediction model of a long short-term memory network algorithm based on the cleaning state label, input historical cleaning state data, perform trend prediction, and obtain a cleaning state prediction result within a set time range in the future; select a chart type based on the cleaning state prediction result, the statistical indicators of the key parameters and the cleaning state label, the chart type including a line chart and a bar chart, and obtain visual graphic data; construct a graphical user interface based on the visual graphic data, integrate the visual graphics, perform interface layout design, and obtain a final visual monitoring platform interface.

[0046] Specifically, assume that the vibration, temperature and pressure data of the cleaning equipment are collected through sensors, with 100 vibration data points per second, 1 temperature data point per second, and 5 pressure data points per second, and the collection lasts for 10 minutes. The original data set contains 60,000 vibration data points (for example: 12g, 15g, 13g...), 600 temperature data points (for example: 25℃, 26℃, 27℃...), and 3,000 pressure data points (for example: 52MPa, 53MPa, 51MPa...). Arrange these data in chronological order. Due to the different data collection frequencies, data preprocessing is performed to uniformly convert all data into 1 data point per second. The average method is used, for example, averaging every 100 vibration data points to obtain 600 vibration data points, and averaging every 5 pressure data points to obtain 600 pressure data points. In this way, 600 vibration data points, 600 temperature data points and 600 pressure data points are obtained, and the timestamps of all data are aligned. Then, the statistical process control (SPC) algorithm is used to calculate the mean and standard deviation of these 600 data, and the vibration mean is 14g, the standard deviation is 0.2g; the temperature mean is 20℃, the standard deviation is 5℃; the pressure mean is 52MPa, and the standard deviation is 0.1MPa. These statistical indicator data are input into the pre-trained random forest model, which is trained based on historical data and contains 100 decision trees, each with a maximum depth of 10. By voting for each input statistical indicator, the cleaning status label is output, such as "normal", "slightly dirty", and "severely dirty". Assume that the current input data is labeled as "slightly dirty". And according to the proportion of the three states of "normal", "slightly dirty", and "severely dirty" in the historical data, for example, the data of the past month shows that the normal state accounts for 70%, the light dirt accounts for 20%, and the severe dirt accounts for 10%. The cleaning status labels of every 10 minutes in the past month are used as input, and the Long Short-Term Memory Network (LSTM) model is used for prediction. The model consists of an input layer, two hidden layers (50 neurons in each layer), and an output layer. By learning the changing trend of the historical status, the cleaning status of the next hour is predicted. For example, the prediction results show that in the next hour, the first 30 minutes are "slightly dirty", and there is a 70% probability of turning into "severely dirty" in the next 30 minutes. The current key parameters (vibration, temperature, pressure), cleaning status labels (slightly dirty), and prediction results (status changes in the next hour) are integrated, and a line chart is selected to show the changing trend of vibration, temperature, and pressure, and a bar chart of different colors is used to represent the current and predicted cleaning status. Finally, these charts are integrated into a graphical user interface, and the interface layout is designed, such as placing the real-time data chart on the left, the prediction results on the right, and adding equipment information and control buttons on the top. Users can use this interface to monitor the status of the cleaning equipment in real time and view the prediction results of the future status.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time detection and monitoring of semiconductor wafer cleaning effect, characterized in that: The method comprises the following steps: Step S101, using high-sensitivity nano-scale particle detection technology and optical sensing equipment to capture the scattered signals of sub-nanometer particles on the wafer surface, and improving the sensitivity and accuracy of particle detection through signal amplification and noise suppression algorithms; Step S102: Introduce rapid chemical analysis technology, combine electrochemical sensors and spectral analysis, monitor the metal ion residues on the wafer surface in real time, shorten the detection cycle, and achieve instant feedback on the production line; Step S103, design an intelligent parameter monitoring system to collect parameters of the composition, concentration, temperature and cleaning time of the cleaning solution in real time, use a machine learning algorithm to analyze the impact of parameter fluctuations on the cleaning effect, and dynamically adjust the cleaning process parameters to ensure a stable cleaning effect; Step S104: Develop a multimodal data fusion algorithm to integrate optical detection data, chemical analysis data, and cleaning parameter data, build a comprehensive evaluation model, comprehensively evaluate the wafer cleaning effect, and improve the reliability of the detection results; Step S105: applying deep learning image processing technology to denoise, enhance and extract features from the captured wafer surface image, accurately identify and quantify nano-scale particles, and generate a high-resolution surface cleanliness distribution map; Step S106: construct a real-time feedback control system to automatically adjust the working parameters of the cleaning equipment according to the output of the comprehensive evaluation model, optimize the cleaning process, and reduce the amount of cleaning fluid used and the environmental impact; Step S107: design a data-driven cleaning effect prediction model, use historical cleaning data and real-time monitoring data to predict future cleaning effect trends, warn of potential problems in advance, and improve the preventive maintenance capabilities of the production line; Step S108: Integrate edge computing technology to transfer data processing and analysis tasks to the cleaning equipment end, reduce data transmission delays, and improve the response speed of real-time monitoring and feedback control; Step S109: Develop a visual monitoring platform to display key parameters, test results and predicted trends of the cleaning process in a graphical manner, so that operators can understand the cleaning status in real time and respond quickly to abnormal situations.

2. The real-time detection and monitoring method for semiconductor wafer cleaning effect according to claim 1, characterized in that: The step S101 comprises: Acquire the scattering signal on the wafer surface; If the light intensity distribution of the scattered signal conforms to the Gaussian distribution, amplifying the scattered signal to obtain an amplified scattered signal; If the light intensity distribution of the scattered signal does not conform to the Gaussian distribution, the scattered signal is not amplified; Performing Fourier transform on the amplified scattered signal to obtain frequency domain data; If the proportion of high-frequency components in the frequency domain data exceeds a preset threshold, Gaussian filtering is performed on the frequency domain data to obtain denoised frequency domain data; If the proportion of high-frequency components in the frequency domain data does not exceed a preset threshold, no Gaussian filtering is performed on the frequency domain data; Performing inverse Fourier transform on the denoised frequency domain data to obtain a time domain signal; Determining the peak intensity and the number of peaks of the scattered signal according to the peak characteristics of the time domain signal; The size and quantity of the sub-nano particles on the wafer surface are determined according to the peak intensity and the peak quantity of the scattering signal.

3. The real-time detection and monitoring method for semiconductor wafer cleaning effect according to claim 1, characterized in that: The step S102 comprises: Acquire a set of original electrical signals from the surface layer of the wafer, and perform Fourier transform on the original electrical signals to obtain frequency domain data of the electrical signals; Acquiring spectral data of the surface layer of the wafer, and calculating the spectral absorptivity by least square method based on the spectral data; Determine the type of residue information by comparing the spectral absorbance with a pre-established metal ion bond absorbance database; Acquiring the residue type information and the electrical signal frequency domain data, and obtaining fused data according to the residue type information and the electrical signal frequency domain data; According to the fused data, the data is input into a trained artificial neural network model to obtain a predicted value of the metal ion concentration in the surface layer; According to the comparison between the predicted value of the metal ion concentration of the surface layer and a preset threshold value, if the predicted value is greater than the threshold value, an alarm signal is generated; Acquire the alarm signal, and control the production line to suspend operation according to the alarm signal; A visual report is generated according to the alarm signal, wherein the report includes information on the type of residue, the concentration of residue, and the location of residue.

4. The real-time detection and monitoring method for semiconductor wafer cleaning effect according to any one of claims 1 to 3, characterized in that: The step S104 comprises: Obtaining image information output by an optical detection device, extracting scattering state information on the wafer surface using an image processing algorithm, and obtaining a first scattering feature vector; Obtaining a first scattering characteristic vector and a metal residue amount on a wafer surface output by a chemical analyzer, and performing normalization processing using a data normalization algorithm to obtain first fused data; According to the first fused data and the cleaning pressure value output by the cleaning equipment, a principal component analysis algorithm is used to perform dimensionality reduction to obtain second fused data; According to the second fusion data and the cleaning temperature value output by the cleaning equipment, a random forest regression model is constructed to obtain a predicted value of wafer cleanliness; Obtain the predicted value of wafer cleanliness and the type of cleaning solvent output by the cleaning equipment, build a support vector machine classification model, and obtain the wafer cleaning effect level; According to the wafer cleaning effect level and the cleaning time output by the cleaning equipment, an XGBoost regression model is constructed to obtain the optimized cleaning time; A closed-loop feedback control system is constructed based on the optimized cleaning time and the number of particles on the wafer surface output by the optical inspection equipment. If the number of particles on the wafer surface is greater than the preset threshold, the cleaning time is adjusted. If the number of particles on the wafer surface is not greater than the preset threshold, the cleaning process is determined to be completed.

5. The real-time detection and monitoring method for semiconductor wafer cleaning effect according to any one of claims 1 to 3, characterized in that: The step S105 comprises: Collecting original images of the wafer surface; Using a pre-established convolutional neural network model, a convolution operation is performed on the original image to obtain a first image; According to the noise level of the first image, setting Gaussian filter parameters; Using the Gaussian filter parameters, performing a Gaussian filter operation on the first image to obtain a denoised second image; According to the contrast of the second image, a histogram equalization algorithm is used to perform grayscale mapping on the second image to determine an enhanced third image; Acquiring gradient information of the third image; Using a preset Canny edge detection operator to perform edge detection processing on the third image to obtain a fourth image, wherein the fourth image contains edge information; According to the edge information of the fourth image, a morphological processing method is selected to perform particle segmentation to obtain a segmented fifth image, wherein the fifth image contains particle contour information; For the particle contour information contained in the fifth image, the particle size distribution data is obtained by calculating the pixel area of ​​each particle contour region; Calculate the number of pixels of each particle to obtain the particle quantity data; The particle size distribution data and the particle quantity data are mapped to corresponding spatial coordinates to generate a surface cleanliness distribution map.

6. The real-time detection and monitoring method for semiconductor wafer cleaning effect according to any one of claims 1 to 3, characterized in that: The step S106 comprises the following steps: Step 1: Collect image information of the surface of the cleaning object, extract image feature information through image recognition algorithm, and obtain the initial cleanliness value of the cleaning object; Step 2: According to the initial cleanliness value of the cleaning object, the pre-established database of cleaning liquid concentration, temperature and cleaning time is matched to determine the initial working parameter combination of the cleaning equipment; Step 3: The cleaning equipment is operated under the initial working parameter combination, and the sensor monitors the turbidity and conductivity index data of the discharged waste liquid during the cleaning process to obtain real-time data; Step 4: Compare the real-time data with the historical data corresponding to the initial cleanliness value of the cleaning object, the cleaning solution concentration, temperature and cleaning time data. If the deviation between the real-time data and the historical data exceeds a preset threshold, the adjustment strategy is triggered; Step 5: Adjust the strategy based on the deviation direction and magnitude between real-time data and historical data, and use the long short-term memory network algorithm to predict the change trend of the cleanliness of the cleaning object and obtain the predicted final cleanliness value; Step 6: Determine whether the predicted final cleanliness value matches the preset cleanliness target value. If not, use the gradient boosting decision tree algorithm to output the equipment parameter adjustment instruction and obtain the adjustment amount of the equipment parameter. Step 7: The control system receives the adjustment amount of the equipment parameters, adjusts the cleaning liquid concentration, temperature and cleaning time parameters of the cleaning equipment, and executes step 3 repeatedly until the predicted final cleanliness value matches the preset cleanliness target value to obtain an optimized cleaning process.

7. The real-time detection and monitoring method for semiconductor wafer cleaning effect according to any one of claims 1 to 3, characterized in that: The step S109 comprises: Acquiring equipment operation data, wherein the equipment operation data includes vibration data, temperature data, and pressure data collected by sensors; According to the device operation data, data format conversion is performed to obtain a pre-processed data set in a unified data format; Calculating statistical indicators of key parameters according to the preprocessed data set, wherein the statistical indicators include mean and standard deviation; According to the statistical indicators of the key parameters, a pre-established random forest algorithm model is applied to perform cleaning state identification to obtain a cleaning state label corresponding to each data in the pre-processed data set; According to the cleaning status label, a prediction model of a long short-term memory network algorithm is constructed, historical cleaning status data is input, trend prediction is performed, and a cleaning status prediction result within a future set time range is obtained; According to the cleaning state prediction result, the statistical index of the key parameter and the cleaning state label, a chart type is selected, and the chart type includes a line chart and a bar chart to obtain visual graphic data; According to the visualization graphic data, a graphical user interface is constructed, the visualization graphics are integrated, and the interface layout is designed to obtain the final visualization monitoring platform interface.