Chip appearance defect identification method based on improved random forest algorithm
By improving the random forest algorithm, streamlining the decision tree structure, optimizing the decision tree weights and automatic optimization parameters, the problems of low accuracy and low efficiency of chip appearance defect recognition are solved, and more efficient and reliable defect recognition effects are achieved.
Patent Information
- Application Number
- CN202411436893.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2025-02-25
AI Technical Summary
In the prior art, chip appearance defect recognition has low accuracy, low efficiency and limited ability to detect complex defects.
The improved random forest algorithm is used to streamline the decision tree structure through mutual information technology to remove redundant decision trees with low classification performance; the genetic algorithm is used to optimize the decision tree weights to reduce the risk of overfitting; the grid search method is used to automatically optimize the key parameters of the random forest to improve the overall performance of the model.
It improves the efficiency and accuracy of the model, enhances the generalization ability and robustness of the model, simplifies the parameter configuration process, and ensures the stability and reliability of the model in practical applications.
Smart Images
Figure CN119515778A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of semiconductor manufacturing and detection, and in particular relates to a chip appearance defect recognition method based on an improved random forest algorithm. Background Art
[0002] In the semiconductor manufacturing industry, the quality of chips is directly related to the performance and reliability of the final product. With the continuous advancement of integrated circuit technology, chip design has become increasingly complex, which not only increases the difficulty of the manufacturing process, but also increases the possibility of potential defects. Efficient and accurate chip defect detection technology has become one of the key factors to ensure product quality, improve production efficiency and reduce production costs. Traditional manual inspection methods have been difficult to adapt to the rhythm of modern large-scale production. Its main limitations are slow speed, high labor costs and susceptibility to human factors. In order to overcome these limitations, automated inspection technology has emerged. With the help of advanced imaging technology and computer vision algorithms, automated inspection systems can identify and locate defects on chips at a faster speed and higher accuracy, thereby greatly improving the efficiency and quality control level of the entire manufacturing process. This automated solution can not only significantly reduce errors and omissions, but also provide real-time feedback on inspection results, provide instant data support for the production line, and then optimize the production process and improve the yield rate. It is an indispensable and important part of the semiconductor manufacturing industry.
[0003] Among many machine learning algorithms, random forests have performed well in the fields of image classification and defect detection due to their strong generalization ability, robustness, and efficient computing performance. Random forests effectively reduce the risk of overfitting and enhance the stability and accuracy of the model by constructing multiple decision trees and integrating their prediction results. This feature is particularly important for processing high-dimensional data, because image data usually contains a large number of pixels as features, and random forests can automatically select key features to avoid unnecessary computational overhead. In addition, random forests have parallel processing capabilities and can fully utilize the advantages of multi-core processors to process large amounts of data in a short time, which is very important for real-time or large-scale image data sets. However, the general random forest algorithm also has shortcomings in image defect recognition tasks. First, it may introduce redundant decision trees, increase the computational burden, and reduce model efficiency and accuracy. Second, there may be a high correlation between randomly generated decision trees, which affects model diversity and weakens overall performance. In addition, the parameter configuration of random forests is relatively complex, and manual adjustment to obtain optimal performance is a time-consuming and difficult task. Therefore, it is necessary to streamline the structure of the random forest algorithm to improve its performance in chip appearance defect recognition. Summary of the invention
[0004] The purpose of the present invention is to provide a chip appearance defect recognition method based on an improved random forest algorithm to solve the problems of low accuracy, low efficiency and limited ability to detect complex defects in chip appearance defect recognition in the prior art.
[0005] To achieve the above object, the present invention adopts a chip appearance defect recognition method based on an improved random forest algorithm, which specifically includes the following steps:
[0006] Step 1: Data collection: The acquisition system collects chip image data containing various types of defects;
[0007] Step 2: Image preprocessing: Perform preprocessing operations on the acquired images to improve image quality and obtain more obvious defect features to be extracted;
[0008] Step 3: Feature extraction and optimization: Use mutual information technology to streamline the decision trees in the random forest and remove redundant decision trees with low classification performance, thereby improving the efficiency and accuracy of the model;
[0009] Then, the weights of the streamlined decision tree are optimized through genetic algorithms to reduce the risk of overfitting and enhance the generalization ability of the model;
[0010] Step 4: Model training: Use the grid search method to further optimize the key parameters of the random forest (including the number of trees, maximum depth, etc.) to improve the overall performance of the model;
[0011] Step 5: Performance evaluation and optimization: Evaluate the performance of the model through cross-validation method to ensure the stability and reliability of the model in practical applications;
[0012] If the detection performance meets the preset detection requirements, the training is terminated; otherwise, the model training step is returned to adjust the model parameters and retrain.
[0013] Preferably, in step 2, the image preprocessing method includes performing image preprocessing operations such as noise reduction, enhancement, and segmentation on the collected image, so as to obtain a region of interest or an input size more suitable for the image detection algorithm, and extract the features that are most likely to distinguish defects from the background. At the same time, the image quality is improved to ensure the effectiveness of subsequent analysis.
[0014] Preferably, in step 3, the mutual information technology is used to calculate the correlation between decision trees, remove redundant decision trees with low classification performance, and thus streamline the random forest.
[0015] Preferably, in step 3, the genetic algorithm is used to select the most discriminative features from the preprocessed image to improve the accuracy and stability of the detection model. Specifically, the optimal feature combination is determined through the iterative process of the genetic algorithm to optimize the model performance. Genetic operations such as selection, crossover, and mutation are used to gradually screen out the feature subset that is most conducive to the detection task.
[0016] Preferably, in step 4, the grid search method is used to determine the optimal parameter configuration in the random forest algorithm, including but not limited to the number of trees, the maximum depth of the trees, etc., to optimize the model performance.
[0017] Preferably, in step 5, during the performance evaluation and optimization process, cross-validation is used to evaluate the model performance under different parameter configurations, so as to select the optimal configuration.
[0018] Compared with the prior art, the advantages of the present invention are:
[0019] 1. Streamline the decision tree structure to improve model efficiency and accuracy.
[0020] The present invention uses mutual information technology to calculate the correlation between decision trees, removes redundant decision trees with low classification performance, and thus simplifies the random forest. This process reduces the computational burden and improves the operating efficiency and accuracy of the model.
[0021] 2. Optimize decision tree weights and enhance model generalization capabilities.
[0022] The present invention introduces a genetic algorithm to optimize the weights of the streamlined decision tree, and selects the feature subset that is most conducive to the detection task through operations such as selection, crossover, and mutation. This method effectively reduces the risk of overfitting and enhances the generalization ability and robustness of the model.
[0023] 3. Automatic parameter optimization to simplify the configuration process.
[0024] The present invention uses a grid search method to optimize the key parameters of the random forest (such as the number of trees, maximum depth, etc.), and evaluates the model performance under different parameter configurations through cross-validation to select the optimal configuration. This simplifies the parameter configuration process and ensures the optimal performance of the model in various situations.
[0025] 4. Comprehensive performance evaluation to ensure stability and reliability.
[0026] The present invention uses a cross-validation method to comprehensively evaluate the model performance to ensure the stability and reliability of the model in practical applications. If the detection performance does not meet the preset requirements, the system will automatically return to the training step, adjust the parameters and retrain until a satisfactory performance level is achieved.
[0027] In summary, the present invention not only improves the efficiency and accuracy of the random forest model, but also enhances the generalization ability and stability of the model by streamlining the decision tree structure through mutual information technology, optimizing the decision tree weight through genetic algorithm, and automatically optimizing the parameters through grid search method. These improvements provide a more efficient and reliable solution for chip appearance defect recognition, and are particularly suitable for the automated inspection system in the semiconductor manufacturing industry. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 It is the overall flow chart of applying the improved random forest algorithm to chip appearance defect recognition.
[0029] Figure 2 It is a schematic diagram of the simplified random forest structure of mutual information technology.
[0030] Figure 3 It is a schematic diagram of optimizing decision tree feature weights based on genetic algorithm. DETAILED DESCRIPTION
[0031] The following is a more detailed description of the chip appearance defect recognition method based on the improved random forest algorithm of the present invention in conjunction with a schematic diagram, wherein a preferred embodiment of the present invention is shown, and it should be understood that those skilled in the art can modify the present invention described herein while still achieving the advantageous effects of the present invention. Therefore, the following description should be understood as being widely known to those skilled in the art, and not as a limitation of the present invention.
[0032] Figure 1 This is the overall flow chart of applying the improved random forest algorithm to chip appearance defect recognition. As shown in the figure, the specific steps are as follows:
[0033] Step 1: Data acquisition: The acquisition system collects chip image data containing various types of defects.
[0034] Step 2: Image preprocessing: Perform preprocessing operations on the acquired images to improve image quality and obtain more obvious defect features to be extracted.
[0035] Among them, the image preprocessing method includes image preprocessing operations such as noise reduction, enhancement, and segmentation on the collected images, so as to obtain the region of interest or the input size more suitable for the image detection algorithm, and extract the features that are most likely to distinguish defects from the background. At the same time, it improves the image quality and ensures the effectiveness of subsequent analysis.
[0036] Step 3: Feature extraction and optimization: Use mutual information technology to streamline the decision trees in the random forest and remove redundant decision trees with low classification performance, thereby improving the efficiency and accuracy of the model.
[0037] Then, the weights of the streamlined decision tree are optimized through genetic algorithm to reduce the risk of overfitting and enhance the generalization ability of the model.
[0038] Specifically, mutual information technology is used to calculate the correlation between decision trees and remove redundant decision trees with low classification performance, thereby streamlining the random forest.
[0039] Then, a genetic algorithm is used to select the most discriminative features from the preprocessed images to improve the accuracy and stability of the detection model.
[0040] Specifically, the best feature combination is determined through the iterative process of the genetic algorithm to optimize the model performance. Genetic operations such as selection, crossover, and mutation are used to gradually select the feature subset that is most beneficial to the detection task.
[0041] Step 4: Model training: Use the grid search method to further optimize the key parameters of the random forest (including the number of trees, maximum depth, etc.) to improve the overall performance of the model.
[0042] Specifically, the grid search method is used to determine the best parameter configuration in the random forest algorithm, including but not limited to the number of trees, the maximum depth of the trees, etc., to optimize the model performance.
[0043] Step 5: Performance evaluation and optimization: Evaluate the performance of the model through cross-validation method to ensure the stability and reliability of the model in practical applications.
[0044] If the detection performance meets the preset detection requirements, the training is terminated; otherwise, the model training step is returned to adjust the model parameters and retrain.
[0045] Figure 2 A schematic diagram of the simplified random forest structure of the mutual information technology. As shown in the figure, the random forest algorithm uses multiple decision trees to train and predict chip defect samples.
[0046] Assume that the original data after preprocessing is:
[0047] H={(x1,y1),(x2,y2),...,(x i ,y i ),...,(x M ,y M )} (1)
[0048] Among them, x i is the original chip defect data, with dimension N; y i is the defect type; M is the number of sample data.
[0049] When training the random forest defect recognition model, this paper first uses the Bootstrap method to randomly extract n samples from the chip image dataset H with replacement and constructs n decision trees.
[0050] Each tree is then split according to the Gini coefficient until it cannot be split any more.
[0051] The Gini coefficient is defined as:
[0052]
[0053] Among them, p is the probability distribution of the data set, and s is the number of categories in the data set.
[0054] Repeat the above process to generate multiple classification trees and establish an initial random forest model. When using the random forest model to classify and predict chip image data, each decision tree in the random forest acts as a predictor to classify the chip image data, and then the final defect recognition result is obtained by aggregating the classification results of all decision trees in the random forest.
[0055] Then, the random forest structure is streamlined using mutual information technology.
[0056] Specifically, for each pair of decision trees, the mutual information between them is calculated, which measures the correlation between the two trees in the prediction results. The higher the mutual information value, the stronger the correlation between the two trees and the higher the redundancy.
[0057]
[0058] Among them, T i and T j Represents two decision trees, X and Y correspond to the set of prediction results of the two trees, p(x,y), p(x) and p(y) represent the joint probability distribution and marginal probability distribution respectively.
[0059] Based on the calculated mutual information values, a correlation matrix is constructed. Each element in the matrix represents the mutual information value between the corresponding decision tree pairs.
[0060] Then set the mutual information threshold θ to determine whether the correlation between decision trees is too high. From the correlation matrix, find the decision tree pairs whose mutual information value is higher than the threshold θ. Keep one of the decision trees with higher classification performance and remove the other redundant decision tree.
[0061] Preferably, the decision of which tree to keep is made by evaluating the F1 score indicator of each tree. After removing the redundant decision trees, the random forest structure is updated to form a new streamlined random forest.
[0062] Figure 3This is a schematic diagram of optimizing the feature weights of a decision tree based on a genetic algorithm. As shown in the figure, it mainly includes the following steps:
[0063] 1. Initial population generation
[0064] After the simplified random forest is obtained based on the mutual information calculation, there are n new A decision tree.
[0065] In the genetic algorithm, each individual represents a set of decision tree weight configurations, so each individual can be represented as an n new dimensional vector, where each element corresponds to the weight w of a decision tree.
[0066] 2. Weight initialization
[0067] Preferably, a random initialization method is used. The initial weight of each decision tree is It can be randomly generated, randomly drawing weights from a uniform distribution [0,1] and normalizing them so that the sum of all weights is 1.
[0068]
[0069] 3. Calculate fitness
[0070] For each individual, a weighted vote is performed on the random forest model using its corresponding weight configuration, and then the performance of the model is evaluated on the validation set to obtain the fitness value of the individual.
[0071] 4. Selection, crossover and mutation operations
[0072] According to the fitness value, the individuals with better performance are selected from the current population as the parent generation. Preferably, a roulette selection strategy is adopted, in which individuals with higher fitness have a greater probability of being selected as the parent generation. The selected parent generation individuals will be used to generate the next generation population.
[0073] Next, the selected parent individuals are paired up and new offspring individuals are generated through crossover operations. The newly generated offspring inherit the characteristics from different parents, thus generating new weight configurations.
[0074] Common crossover methods include single-point crossover, multi-point crossover, or uniform crossover, etc. For example, in single-point crossover, a random point is selected to split the weight vectors of two parent individuals into two parts, and these parts are exchanged to generate two new offspring individuals.
[0075] In order to maintain population diversity and explore the solution space, mutation operations are performed on a portion of the offspring individuals. Mutation can be achieved by randomly changing the weights of certain decision trees, such as randomly perturbing a weight value within a small range. Specifically, some offspring individuals can be selected, and one or more elements in their weight vectors can be randomly selected and randomly adjusted within a certain range. After mutation, the weight configuration of some offspring individuals will change, introducing new solutions, which helps prevent premature convergence to the local optimal solution and increases the possibility of finding the global optimal solution.
[0076] 5. New population generation
[0077] After selection, crossover and mutation, a new generation of population is formed. This new population contains a new weight configuration and continues the next round of fitness evaluation and optimization.
[0078] 6. Set termination conditions
[0079] The above steps (fitness evaluation, selection, crossover, mutation) are repeated until the preset termination condition is met. Specifically, a maximum number of iterations Gen is set. If the current generation gen reaches the maximum value Gen, the optimization is stopped; if the fitness value Fitness reaches the preset threshold ε, the optimization can also be terminated early. After multiple iterations of optimization, a set of optimized decision tree weight configurations are finally obtained.
[0080] The embodiments described above are merely descriptions of the present invention and do not limit the scope of the present invention. Those skilled in the art may make various improvements and modifications to the present invention without departing from the design spirit of the present invention, but these improvements and modifications should all fall within the scope of protection determined by the claims of the present invention.
[0081] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.
Claims
1. A chip appearance defect recognition method based on an improved random forest algorithm, characterized in that: The following steps are involved: Step 1: Data collection: The acquisition system collects chip image data containing various types of defects; Step 2: Image preprocessing; Step 3: Feature extraction and optimization: Streamlining decision trees in random forests through mutual information technology; Then, the weights of the streamlined decision tree are optimized through genetic algorithm; Step 4: Model training: Use grid search method to further optimize the key parameters of random forest; Step 5: Performance evaluation and optimization: Evaluate the performance of the model through cross-validation method to ensure the stability and reliability of the model in practical applications; If the detection performance meets the preset detection requirements, the training is terminated; otherwise, the model training step is returned to adjust the model parameters and retrain.
2. The chip appearance defect recognition method based on the improved random forest algorithm according to claim 1, characterized in that: Perform image preprocessing operations such as denoising, enhancement, and segmentation on the collected images.
3. The chip appearance defect recognition method based on the improved random forest algorithm according to claim 1, characterized in that: The mutual information technology is used to calculate the correlation between decision trees and remove redundant decision trees with low classification performance.
4. The chip appearance defect recognition method based on the improved random forest algorithm according to claim 1, characterized in that: The best feature combination is determined through an iterative process of genetic algorithm to optimize model performance.
5. The chip appearance defect recognition method based on the improved random forest algorithm according to claim 1, characterized in that: The grid search method is used to determine the best parameter configuration in the random forest algorithm, including the number of trees and the maximum depth of the trees, to optimize the model performance.
6. According to the chip appearance defect recognition method based on the improved random forest algorithm described in claim 5, cross-validation is used to evaluate the model performance under different parameter configurations, so as to select the optimal configuration.