TFT conductive particle screening method based on multidimensional data analysis

Abstract

A TFT conductive particle screening method based on multidimensional data analysis relates to the field of semiconductor material detection technology. The present invention is intended to solve the problems of low efficiency and poor accuracy in existing methods for detecting TFT conductive particles. The TFT conductive particle screening method based on multidimensional data analysis described in the present invention inputs the multidimensional data of the TFT conductive particles to be screened into a trained multidimensional data analysis model to obtain screening results. The multidimensional data analysis model integrates convolutional neural networks and recurrent neural networks. The former performs feature extraction on particle image data, and the latter performs sequence modeling on the fused multidimensional data. The model can quickly and accurately identify TFT conductive particles that meet specific standards, greatly improving screening efficiency and accuracy, and meeting the needs of large-scale production for rapid and accurate screening of conductive particles.

Description

Technical Field

[0001] The invention belongs to the technical field of semiconductor material detection. Background Art

[0002] With the development of the semiconductor industry, TFT (thin-film transistor) conductive particles are increasingly used in various electronic products, and their quality directly affects product performance.

[0003] At present, the detection of TFT conductive particles mainly relies on a relatively simple detection method. For particle size detection, ordinary optical microscopes are often used to obtain particle size data through manual measurement. This is not only inefficient, but the measurement accuracy is easily affected by human factors. In terms of shape detection, it is difficult to obtain comprehensive three-dimensional shape information of particles by simply relying on two-dimensional image observation under a microscope. In terms of electrical performance testing, simple resistance measuring devices are mostly used, which cannot accurately measure key parameters such as resistivity and conductivity, and the measurement error is large. Chemical composition analysis uses conventional chemical analysis methods, which are cumbersome and time-consuming, and cannot meet the needs of batch testing.

[0004] In summary, existing detection methods for TFT conductive particles have problems such as low efficiency and poor accuracy, which makes it difficult to meet the needs of rapid and accurate screening of conductive particles in large-scale production. Summary of the Invention

[0005] The present invention aims to solve the problems of low efficiency and poor accuracy in existing methods for detecting conductive particles in TFTs, and provides a TFT conductive particle screening method based on multidimensional data analysis.

[0006] A TFT conductive particle screening method based on multidimensional data analysis includes: inputting multidimensional data of TFT conductive particles to be screened into a trained multidimensional data analysis model to obtain screening results.

[0007] Furthermore, the training process of the above-trained multidimensional data analysis model includes:

[0008] Using the multi-dimensional data of the TFT conductive particle sample as a training set, and training the multi-dimensional data analysis model based on a machine learning algorithm;

[0009] The multi-dimensional data includes: size information, shape information, electrical performance information and chemical composition information of TFT conductive particles;

[0010] The machine learning algorithm uses a convolutional neural network to extract size and shape features from image data of TFT conductive particles, then fuses the size and shape features with the electrical performance and chemical composition information of the TFT conductive particles. The fused results are then sequence-modeled using a recurrent neural network to obtain a multidimensional data analysis model capable of identifying TFT conductive particles that meet specific standards.

[0011] Furthermore, the method for obtaining the multi-dimensional data includes:

[0012] Use high-precision electron microscope to obtain the size information of TFT conductive particles;

[0013] The two-dimensional profile of the conductive particles in the TFT is obtained using a scanning electron microscope combined with energy spectrum analysis, and a three-dimensional shape model is generated through a reconstruction algorithm to obtain shape information.

[0014] The electrical properties of TFT conductive particles are measured using a four-probe tester;

[0015] The chemical composition information of TFT conductive particles is obtained by inverting the element content through the intensity of X-ray characteristic peaks.

[0016] Furthermore, the size information of the TFT conductive particles includes the particle size and particle size distribution of the TFT conductive particles. The particle size distribution characteristics are described using the skewness coefficient and the kurtosis coefficient:

[0017]

[0018] Among them, x i is the particle size measurement value of the i-th TFT conductive particle, x is the average particle size, σ is the standard deviation of the particle size, n is the total number of TFT conductive particles, and i = 1, 2, ..., n.

[0019] Furthermore, the electrical performance information includes resistivity ρ and conductivity

[0020]

[0021] Wherein, R is the measured resistance value, A is the effective cross-sectional area of ​​the current path of the TFT conductive particles when measuring its electrical properties, and L is the probe spacing.

[0022] Furthermore, after obtaining the multi-dimensional data, the multi-dimensional data is pre-processed, and all the pre-processed data is used as a sample set;

[0023] The preprocessing includes data cleaning and filling in missing data.

[0024] Furthermore, the above data cleaning includes:

[0025] Initial screening: Use the mean and standard deviation of multi-dimensional data to construct a data sample range, and delete multi-dimensional data outside the sample range. The data sample range is:

[0026]

[0027] Among them, α is the data sample feature, is the mean of α, σ is the standard deviation of α, and k is the threshold coefficient;

[0028] Fine screening: Use the isolation forest algorithm to randomly divide the feature space to generate a segmentation tree, analyze the depth required for the remaining multi-dimensional data to be isolated in the tree, and calculate the anomaly score s(y) of each data sample y in the remaining multi-dimensional data:

[0029]

[0030] Among them, h(y) is the average isolation depth of data sample y in the isolation forest, c(n) is the adjustment factor for the average path length of trees in the forest,

[0031] If s(y) is close to 1, the data sample y is close to an outlier; if s(y) is close to 0, the data sample y is close to a normal value. The remaining multidimensional data is filtered according to s(y).

[0032] Furthermore, the above-mentioned filling of missing data includes:

[0033] When the missing data is numerical data, if the distribution of the missing data is close to the normal distribution, the mean or median of the data adjacent to the missing data will be directly filled in. If the distribution of the missing data is complex and cannot be modeled, the K nearest neighbor algorithm will be used to fill in the missing values.

[0034] When the missing data is non-numeric data, the decision tree model is used to predict the missing values.

[0035] Furthermore, the above K-nearest neighbor algorithm calculates the Euclidean distance between sample points, selects K most similar samples, and calculates the filling value v 填补 :

[0036]

[0037] Among them, ν j is the eigenvalue of the jth neighbor sample point, d j is the distance between the current sample point and the jth neighbor sample point.

[0038] Furthermore, a decision tree is used to analyze the information gain between each known data feature and the target category, splitting the data space layer by layer to predict the category to which the missing data belongs;

[0039] The information gain IG(D,A) is:

[0040]

[0041] Among them, H(D) is the information entropy of the current data set D, B is the split feature, and D v It is a subset when the split feature B takes the value V, where V is used to traverse all possible values ​​of the feature.

[0042] The TFT conductive particle screening method based on multidimensional data analysis of the present invention has the following beneficial effects:

[0043] 1. We used a variety of equipment, including high-precision electron microscopes, scanning electron microscopes combined with energy spectrum analysis technology, and four-probe testers, to collaboratively measure the multi-dimensional information of TFT conductive particles, including size, shape, electrical properties, and chemical composition. We collected multiple sets of data for each sample particle. Furthermore, during data preprocessing, we used a combination of statistical methods and machine learning algorithms to remove outliers and fill in missing values ​​using appropriate methods based on data type and distribution. This ensured data integrity, accuracy, and consistency, providing a solid and reliable foundation for subsequent analysis.

[0044] 2. The constructed multidimensional data analysis model integrates convolutional neural networks and recurrent neural networks. The former extracts features from particle image data, while the latter performs sequence modeling on the fused multidimensional data. During the training process, the use of a large amount of labeled sample data and regularization techniques improved the model's accuracy and generalization capabilities. Through cross-validation evaluation and targeted optimization, the model can quickly and accurately identify TFT conductive particles that meet specific standards, significantly improving screening efficiency and accuracy, meeting the needs of large-scale production for rapid and accurate screening of conductive particles.

[0045] 3. After the screening results are output, the original multi-dimensional data of the particles, the screening results, the screening time, and other information are stored in a relational database. A data index is established to facilitate data query and management. For unqualified particles, the reasons for failure are recorded in detail. By analyzing the screening results, product quality issues can be traced, providing data support for optimizing screening methods and improving production processes, thereby improving the quality control level of the entire production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The overall flow chart of the TFT conductive particle screening method based on multi-dimensional data analysis;

[0047] Figure 2 Schematic diagram of a method for obtaining multi-dimensional data of TFT conductive particle samples;

[0048] Figure 3 Remove process views for outliers;

[0049] Figure 4 Fill missing values ​​in the flow view. DETAILED DESCRIPTION

[0050] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other in the absence of conflict.

[0051] Refer to the figure Figures 1 to 4 Specifically describing this embodiment, the TFT conductive particle screening method based on multidimensional data analysis described in this embodiment includes:

[0052] Step 1: Acquire multi-dimensional data of a TFT conductive particle sample, where the multi-dimensional data includes at least: particle size information, shape information, electrical property information, and chemical composition information.

[0053] In this step, particle size information, shape information, electrical properties information, and chemical composition information are measured using high-precision electron microscopes, scanning electron microscopes combined with energy spectrum analysis technology, and four-probe testers to ensure data integrity and high accuracy. The details are as follows:

[0054] (1) Obtain particle size information through high-precision electron microscopy, where particle size and particle size distribution can be accurately quantified using statistical methods. The particle size distribution characteristics are described using the skewness coefficient and the kurtosis coefficient, which are described by the following formulas:

[0055]

[0056] Among them, x i is the particle size measurement value of the i-th particle, x is the average particle size, σ is the standard deviation of the particle size, and n is the total number of particles.

[0057] In statistics, the kurtosis of a normal distribution is defined as 3. The kurtosis coefficient in formula (2) minus 3 is the result that can intuitively represent the deviation of the kurtosis of the data distribution from the normal distribution. The above formula can be used to quantify the symmetry and concentration of the particle size distribution, providing basic data for subsequent particle size classification and screening.

[0058] (2) Using scanning electron microscopy combined with energy spectrum analysis technology, not only can the two-dimensional contours of the particles be obtained, but also a three-dimensional shape model can be generated through a specific reconstruction algorithm to obtain shape information. The shape characteristics are quantitatively described using aspect ratio (AspectRatio) and surface roughness (Surface Roughness). Among them, the aspect ratio is calculated as the ratio of the longest axis length to the shortest axis length of the particle, and the surface roughness is extracted from the surface height fluctuation through Fourier transform to ensure that the shape characteristics of each particle are accurately characterized.

[0059] (3) The electrical properties of the particles are measured by a four-probe tester. The electrical properties include resistivity and conductivity. The resistivity ρ is calculated from the resistance formula Calculation, where R is the measured resistance value, A is the cross-sectional area (the effective cross-sectional area of ​​the current path when the TFT conductive particles are measuring their electrical properties), L is the probe spacing, and conductivity By obtaining multiple sets of data through repeated measurements and combining statistical methods to calculate measurement errors, the reliability of electrical performance data is ensured.

[0060] (4) For chemical composition information, energy spectrum analysis technology is based on high-precision element distribution maps and inverts element content through the intensity of X-ray characteristic peaks. The principal component analysis (PCA) method is combined to reduce the dimensionality of high-dimensional chemical composition data to highlight the main chemical features and provide optimized data for subsequent model training.

[0061] During the data collection process, multiple sets of data are collected for each sample particle. Real-time processing methods such as outlier removal, missing value interpolation, and noise filtering ensure the consistency and accuracy of the resulting multi-dimensional data. This multi-device collaboration and multi-set data collection approach provides a rich and high-quality data foundation for subsequent machine learning-based multi-dimensional data analysis models, making the TFT conductive particle screening process more efficient and accurate.

[0062] Step 2: Preprocess the multidimensional data, including data cleaning, so that outliers and noise data are removed and missing data are filled.

[0063] (1) In the data cleaning process, outliers are removed by combining statistical methods and machine learning algorithms.

[0064] In the outlier identification phase, a preliminary screening mechanism is constructed by calculating the mean and standard deviation of the data to quickly mark potential outliers. Specifically, for each sample feature α, the following formula is used to determine the reasonable range:

[0065]

[0066] in, is the mean of the sample feature α, σ is the standard deviation of the sample feature α, and k is the threshold coefficient.

[0067] To further improve the accuracy of outlier detection, the isolation forest algorithm is used for secondary judgment. By randomly partitioning the feature space to generate a segmentation tree, the depth required for the sample point to be isolated in the tree is analyzed. Outliers are usually isolated with a relatively small number of segmentations. The anomaly score s(y) of sample y is calculated using the following formula:

[0068]

[0069] Here, h(y) represents the average isolation depth of sample y in the isolation forest, and c(n) is an adjustment factor for the average path length of trees in the forest, calculated based on the sample size. If s(y) is close to 1, sample y is very likely an outlier; if s(y) is close to 0, the sample point is more likely a normal value. This combination of statistics and algorithms effectively improves the accuracy and robustness of outlier detection.

[0070] (2) In terms of missing data filling, for numerical data, if its distribution is close to normal distribution, it is directly filled by the mean or median to maintain the stability of the overall data distribution; when the data distribution is complex and cannot be simply modeled, the K nearest neighbor algorithm (KNN) is used to fill the missing values. The KNN algorithm calculates the Euclidean distance between sample points and selects the K most similar samples. Its filling value v 填补 The calculation formula is:

[0071]

[0072] Among them, v i is the eigenvalue of the i-th neighbor sample point, d i is the distance between the current sample point and the i-th neighbor sample point.

[0073] This weighted averaging method ensures that the filling value can reflect the characteristic distribution of the surrounding samples.

[0074] For non-numeric data, such as missing information about certain chemical composition categories of particles, a decision tree model is used to predict missing values. The decision tree analyzes the information gain between each feature in the known data and the target category, splitting the data space layer by layer and ultimately predicting the category to which the missing value belongs. The information gain calculation formula is:

[0075]

[0076] Among them, H(D) is the information entropy of the current data set D, B is the split feature, and D v is the subset when feature B takes the value V. V is used to traverse all possible values ​​of feature B in order to complete the above calculation process, thereby accurately measuring the information gain brought by splitting with feature B, and then finding the best split point, ensuring that the decision tree can accurately predict the category to which the missing value belongs. By maximizing information gain, the decision tree can find the best split point and ensure the accuracy of the prediction.

[0077] Through the above-mentioned sophisticated and targeted data preprocessing method, the quality of the data was effectively improved, laying a solid foundation for the subsequent construction of an accurate multidimensional data analysis model.

[0078] Step 3: Build a multi-dimensional data analysis model based on a machine learning algorithm, which is trained using the pre-processed multi-dimensional data to identify TFT conductive particles that meet specific standards.

[0079] In this step, the machine learning algorithm is a combination model of the convolutional neural network (CNN) and the recurrent neural network (RNN) in the deep learning algorithm. Utilize convolutional neural network to carry out feature extraction to the image data of particles (such as the particle image obtained by microscope, including size and shape information), through structures such as convolutional layer, pooling layer, it is possible to automatically learn the key features in the particle image, such as information such as edge and texture. The extracted image features are fused with the numerical data such as the electrical performance information and chemical composition information of the particle, and then input into the recurrent neural network for further analysis and processing. The recurrent neural network can carry out sequence modeling to the multidimensional data after fusion, fully explore the time series relationship between the data or the potential association between the data, so as to more accurately identify the TFT conductive particles that meet specific standards. In the model training process, a large amount of labeled sample data is used for training, and by continuously adjusting the parameters of the model, the loss function of the model is minimized, and the accuracy and generalization ability of the model are improved. At the same time, regularization techniques such as L1 and L2 regularization are used to prevent model overfitting and ensure that the model can maintain good performance on different data sets.

[0080] Step 4: Input the multi-dimensional data of the TFT conductive particles to be screened into the trained multi-dimensional data analysis model, quickly output the screening results, and determine whether the conductive particles are qualified.

[0081] In this step, before the step of quickly outputting the screening results, the process of evaluating and optimizing the model is also included. Before the multi-dimensional data of the TFT conductive particles to be screened is input into the trained multi-dimensional data analysis model, model evaluation and optimization are key steps to ensure the accuracy of the final screening results. The evaluation process uses the cross-validation method to divide the training data set into k mutually exclusive subsets, select one subset as the validation set in turn, and the remaining subsets as the training set, repeat the model training and validation, and finally take the average result of all rounds as the overall evaluation index of the model. The evaluation indicators of cross-validation include accuracy, recall rate and F1 value, which are defined as follows:

[0082] Accuracy measures the correctness of the model's overall predictions and is calculated as follows:

[0083]

[0084] Among them, TP is the number of true positive cases, TN is the number of true negative cases, FP is the number of false positive cases, and FN is the number of false negative cases.

[0085] In order to more comprehensively measure the balanced performance of the model on positive and negative samples, the F1 value is used as the main optimization target, and its formula is:

[0086]

[0087] in, The F1 value combines Precision and Recall, avoiding the optimization bias that may be caused by a single indicator;

[0088] Based on the aforementioned metrics, the evaluation process analyzes the model's performance across different particle feature dimensions, focusing on identifying data distributions or specific categories with large errors. For areas of poor performance, data preprocessing is first optimized. For example, cluster analysis methods are used to identify the distribution characteristics of particles with unique shapes. The number of samples for these particles is increased to balance the dataset and improve the model's generalization ability. Simultaneously, model hyperparameters are adjusted based on the model's error distribution. The convolutional neural network component optimizes the convolution kernel size to make it more suitable for extracting specific particle shape features. The recurrent neural network component adjusts the number of hidden layer nodes to enhance learning capabilities for time series data (such as changes in particle characteristics over time).

[0089] The rapid screening method also includes recording and managing the screening results. After the screening results are output, the original multi-dimensional data, screening results, screening time and other information of each TFT conductive particle to be screened are stored in the database. The database uses a relational database, such as MySQL. By rationally designing the table structure, the different dimensional data of the particles are stored in different fields, which facilitates data query and management. At the same time, a data index is established to improve the efficiency of data storage and retrieval. For particles that fail the screening results, the reasons for failure are further recorded, such as detailed information such as the size exceeds the standard range and the electrical performance does not meet the standards. The comprehensive recording and management of the screening results not only facilitates the subsequent traceability and analysis of product quality, but also provides data support for further optimization of screening methods and improvement of production processes. For example, by analyzing the data of a large number of unqualified particles, it is found that a certain production link is prone to cause a certain specific defect in the particles, so that targeted adjustments and improvements can be made to this production link.

[0090] Throughout the entire screening process, data security and privacy are strictly protected. During the data collection phase, encryption technology is used to encrypt and transmit the collected raw data, ensuring that it is not stolen or tampered with during transmission. Regarding data storage, data stored in the database is encrypted, using a combination of symmetric and asymmetric encryption algorithms to provide multiple encryption protections for sensitive data. Furthermore, strict access rights are set, ensuring that only authorized personnel can access and manipulate relevant data. For example, data collectors can only read and upload the data they have collected, while data analysts have the authority to analyze and process the data but cannot modify the raw data at will. By establishing a comprehensive data security and privacy protection mechanism, the security and integrity of the multi-dimensional data of TFT conductive particles is ensured throughout the rapid screening process, preventing potential risks to enterprises posed by data leaks.

[0091] Example:

[0092] 10,000 TFT conductive particle samples were randomly selected from the production line. A high-precision electron microscope was used to measure particle size information, obtaining particle size and size distribution data. Statistical calculations were used to determine the skewness and kurtosis coefficients. A scanning electron microscope combined with energy spectrum analysis technology was used to obtain the two-dimensional particle profiles and generate a three-dimensional shape model. Shape features such as aspect ratio and surface roughness were determined. Chemical composition information was also obtained and dimensionality reduction was performed using the PCA method. The resistivity and conductivity of the particles were measured using a four-probe tester. Each group of particles was measured five times, and the measurement error was calculated. During the data collection process, three sets of data were collected for each sample particle, and outliers were removed, missing values ​​were interpolated, and noise was filtered in real time.

[0093] The mean and standard deviation of the data were calculated, and thresholds were set to initially screen out 150 possible outliers. A secondary evaluation using the isolation forest algorithm was then performed, ultimately identifying and removing 120 outliers. For numerical data, such as resistivity data with a normal distribution, the mean was used to fill in five missing values. For shape feature data with complex distributions, the K-nearest neighbor algorithm was used to fill in eight missing values. For the three missing values ​​in chemical composition category information, a decision tree model was constructed to predict and fill in the missing values.

[0094] A combined convolutional neural network and recurrent neural network model was used. Microscope-acquired particle image data was fed into the convolutional neural network. Features were extracted through convolutional and pooling layers, fused with numerical data such as electrical properties and chemical composition, and then fed into the recurrent neural network. The model was trained using 10,000 labeled sample data points and evaluated using cross-validation. Model parameters were adjusted to minimize the loss function, and L1 and L2 regularization was applied to prevent overfitting.

[0095] Using 5-fold cross-validation, the model achieved 96% accuracy, 95% recall, and 95.5% F1 score. Analysis revealed that particles with unusual shapes (aspect ratios greater than 5) exhibited significant prediction errors. Cluster analysis was used to identify the distribution characteristics of these particles. 200 such particle samples were added to balance the dataset. The convolutional neural network kernel size was adjusted to 3×3, and the number of hidden nodes in the recurrent neural network was adjusted to 64. The optimized model's F1 score increased to 97%.

[0096] The optimized model was fed with multi-dimensional data from 1,000 newly produced TFT conductive particles. Within two minutes, the screening process was complete, with 980 particles deemed qualified and 20 deemed unqualified. The original multi-dimensional data, screening results, and screening time for all particles were stored in a MySQL database. The reasons for unqualified particles were recorded, such as five particles with dimensions outside the standard range and 15 particles with substandard conductivity.

[0097] Implementation Effect

[0098] Before implementing this method, traditional detection methods could only detect 1,000-2,000 particles per day, with an accuracy rate of approximately 85%. After adopting this method, the number of particles that can be detected per day increased to 5,000-6,000, with an accuracy rate of 97%. Analysis of the screening results revealed that unstable temperature in a certain production process caused some particles to fail to meet conductivity standards. After adjusting the temperature control parameters in this process, the product failure rate dropped from 5% to less than 2%, effectively improving production efficiency and product quality.

[0099] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely illustrative of the principles and applications of the invention. It should be understood that many modifications may be made to the illustrative embodiments, and that other arrangements may be devised, without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in ways other than those described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in conjunction with other described embodiments.

Claims

1. A TFT conductive particle screening method based on multidimensional data analysis, characterized in that: include: Input the multi-dimensional data of the TFT conductive particles to be screened into the trained multi-dimensional data analysis model to obtain the screening results; The training process of the trained multidimensional data analysis model includes: Using the multi-dimensional data of the TFT conductive particle sample as a training set, and training the multi-dimensional data analysis model based on a machine learning algorithm; The multi-dimensional data includes: size information, shape information, electrical performance information and chemical composition information of TFT conductive particles; The machine learning algorithm uses a convolutional neural network to extract size and shape features from image data of TFT conductive particles, then fuses the size and shape features with electrical performance information and chemical composition information of the TFT conductive particles. The fused results are then sequence-modeled using a recurrent neural network to obtain a multidimensional data analysis model capable of identifying TFT conductive particles that meet specific standards. The method for obtaining multi-dimensional data includes: Use high-precision electron microscope to obtain the size information of TFT conductive particles; The two-dimensional profile of the conductive particles in the TFT is obtained using a scanning electron microscope combined with energy spectrum analysis, and a three-dimensional shape model is generated through a reconstruction algorithm to obtain shape information. The electrical properties of TFT conductive particles are measured using a four-probe tester; The chemical composition information of TFT conductive particles is obtained by inverting the element content through the X-ray characteristic peak intensity; The electrical performance information includes resistivity and conductivity : ， ， in, is the measured resistance value, It is the effective cross-sectional area of ​​the current path of the TFT conductive particles when measuring its electrical properties. is the probe spacing.

2. The TFT conductive particle screening method based on multidimensional data analysis according to claim 1, characterized in that: The size information of the TFT conductive particles includes the particle size and particle size distribution of the TFT conductive particles. The particle size distribution characteristics are characterized by the skewness coefficient. and kurtosis coefficient describe: ， ， in, For the The particle size measurement value of the conductive particles in the TFT, is the average particle size, is the standard deviation of particle size, is the total number of conductive particles in TFT, .

3. The TFT conductive particle screening method based on multidimensional data analysis according to claim 1, characterized in that: After obtaining the multi-dimensional data, the multi-dimensional data is pre-processed, and all the pre-processed data is used as a sample set; The preprocessing includes data cleaning and filling in missing data.

4. The TFT conductive particle screening method based on multidimensional data analysis according to claim 3, characterized in that: Filling in missing data includes: When the missing data is numerical data, if the distribution of the missing data is close to the normal distribution, the missing value is directly filled using the mean or median of the data adjacent to the missing data. If the distribution of the missing data is complex and cannot be modeled, the K-nearest neighbor algorithm is used to fill the missing value. When the missing data is non-numeric data, the decision tree model is used to predict the missing values.

Images