A method and system for generating wafer defect samples
By using radial-based neural networks in wafer defect detection, the problem of insufficient training data is solved and the detection success rate is improved.
Patent Information
- Application Number
- CN202210126893.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-11
- Publication Date
- 2024-12-31
- Estimated Expiration
- 2042-02-11
AI Technical Summary
In the prior art, due to the lack of sufficient training data in wafer defect detection, the training data of deep learning algorithms is insufficient, resulting in a low detection success rate and it is difficult to generate a large number of wafer defect data through small sample data.
By obtaining the labeled wafer defect picture sample set, dividing it into a training set and a test set, a radial-based neural network is established, and the training set is used for model training, the trained network is generated, and the test set is used to perform the test set. Compute, obtain the defect generation training set, rebuild the network, use the defect generation training set for training, and finally use the trained network to automatically generate wafer defect samples.
Generating a large amount of wafer defect data through small sample data solves the problem of insufficient deep learning training data caused by the lack of training sets and improves the success rate of wafer defect detection.
Smart Images
Figure CN114463313B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wafer defect detection, and in particular to a method and system for generating wafer defect samples. Background Art
[0002] Wafer defect detection is an indispensable process in the semiconductor manufacturing and packaging and testing industry. Traditional algorithms usually use differential image comparison to complete defect detection without providing a large number of samples, but traditional methods are easily affected by external interference, resulting in a low detection success rate. After being trained with enough defect samples, the artificial intelligence algorithm based on deep learning has strong generalization ability and strong anti-interference ability. It is currently the most advanced algorithm in the field of wafer defect detection. However, due to the extremely small amount of wafer defect data, it is difficult to obtain the large amount of wafer defect data required for deep learning through large-scale sampling. Therefore, how to generate a large amount of wafer defect data from small sample data has become a research hotspot in recent years. Summary of the invention
[0003] The purpose of the present invention is to provide a method and system for generating wafer defect samples to solve the problem of insufficient deep learning training data caused by the lack of training sets.
[0004] To achieve the above object, the present invention provides the following solutions:
[0005] A method for generating a wafer defect sample, comprising:
[0006] Obtain a marked wafer defect image sample set; the wafer defect image sample set includes multiple wafer defect images; in each of the wafer defect images, a pixel point with a defect is marked as 1, and a pixel point without a defect is marked as 0;
[0007] Dividing the wafer defect image sample set into a training set and a test set;
[0008] Establishing a radial basis function neural network; the input of the radial basis function neural network is the pixel points of the wafer defect image, and the output of the radial basis function neural network is the defect probability;
[0009] Performing model training on the radial basis neural network using the training set to generate a trained radial basis neural network;
[0010] Utilizing the trained radial basis neural network to calculate the test set, to obtain a defect generation training set;
[0011] Using the defect probability as a network input and the pixel defect of the wafer defect image as a network output, reconstructing the radial basis neural network;
[0012] Using the defect generation training set to train the reconstructed radial basis neural network to generate a trained radial basis neural network;
[0013] The trained radial basis function neural network is used to automatically generate wafer defect samples.
[0014] Optionally, dividing the wafer defect image sample set into a training set and a test set specifically includes:
[0015] 50% of the wafer defect images are randomly selected from the wafer defect image sample set as a training set, and the remaining 50% of the wafer defect images are selected as a test set.
[0016] Optionally, the performing model training on the radial basis function neural network by using the training set to generate a trained radial basis function neural network specifically includes:
[0017] The training set is used as training data of the radial basis neural network, the radial basis neural network is trained using a gradient descent method, network parameters are updated, and the trained radial basis neural network is generated.
[0018] Optionally, the using the trained radial basis function neural network to calculate the test set to obtain a defect generation training set specifically includes:
[0019] Calculating the defect probability of each sample in the test set using the trained radial basis neural network;
[0020] The samples with defect probability greater than 50% are classified into the defect generation training set.
[0021] Optionally, the automatically generating wafer defect samples by using the trained radial basis function neural network specifically includes:
[0022] Inputting different probabilities into the trained radial basis neural network for calculation to generate corresponding pixel defects;
[0023] The pixel defect is added to a normal wafer image to generate the wafer defect sample.
[0024] A wafer defect sample generation system, comprising:
[0025] A wafer defect image acquisition module is used to acquire a marked wafer defect image sample set; the wafer defect image sample set includes multiple wafer defect images; a pixel point with a defect in each wafer defect image is marked as 1, and a pixel point without a defect is marked as 0;
[0026] A wafer defect image classification module, used for dividing the wafer defect image sample set into a training set and a test set;
[0027] A radial basis neural network establishment module is used to establish a radial basis neural network; the input of the radial basis neural network is the pixel points of the wafer defect image, and the output of the radial basis neural network is the defect probability;
[0028] A radial basis neural network training module, used for performing model training on the radial basis neural network through the training set to generate a trained radial basis neural network;
[0029] A defect generation training set acquisition module, used to calculate the test set using the trained radial basis neural network to obtain a defect generation training set;
[0030] A radial basis neural network reconstruction module is used to use the defect probability as a network input and the wafer defect image pixel point defect as a network output to reconstruct a radial basis neural network;
[0031] A radial basis neural network training module is reconstructed to train the reconstructed radial basis neural network using the defect generation training set to generate a trained radial basis neural network;
[0032] The wafer defect sample automatic generation module is used to automatically generate wafer defect samples using the trained radial basis function neural network.
[0033] Optionally, the wafer defect image classification module specifically includes:
[0034] The wafer defect image classification unit is used to randomly select 50% of the wafer defect images from the wafer defect image sample set as a training set and the remaining 50% of the wafer defect images as a test set.
[0035] Optionally, the radial basis neural network training module specifically includes:
[0036] The radial basis neural network training unit is used to use the training set as training data of the radial basis neural network, use the gradient descent method to train the radial basis neural network, update the network parameters, and generate the trained radial basis neural network.
[0037] Optionally, the defect generation training set acquisition module specifically includes:
[0038] A defect probability calculation unit, used to calculate the defect probability of each sample in the test set using the trained radial basis neural network;
[0039] The defect generation training set acquisition unit is used to classify the samples with defect probability greater than 50% into the defect generation training set.
[0040] Optionally, the wafer defect sample automatic generation module specifically includes:
[0041] Pixel point defect generation correspondence, used to input different probabilities into the trained radial basis neural network for calculation to generate corresponding pixel point defects;
[0042] The wafer defect sample generating unit is used to add the pixel point defect into the normal wafer image to generate the wafer defect sample.
[0043] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0044] The invention provides a method and system for generating wafer defect samples. The method comprises: obtaining a sample set of wafer defect images after labeling; the sample set of wafer defect images comprises a plurality of wafer defect images; a pixel point with a defect in each of the wafer defect images is labeled as 1, and a pixel point without a defect is labeled as 0; the sample set of wafer defect images is divided into a training set and a test set; a radial basis function neural network is established; the input of the radial basis function neural network is a pixel point of a wafer defect image, and the output of the radial basis function neural network is a defect probability; model training is performed on the radial basis function neural network through the training set to generate a trained radial basis function neural network; the test set is calculated using the trained radial basis function neural network to obtain a defect generation training set; the defect probability is used as a network input, the defect of a pixel point of a wafer defect image is used as a network output, and a radial basis function neural network is reconstructed; the defect generation training set is used to train the reconstructed radial basis function neural network to generate a trained radial basis function neural network; and the trained radial basis function neural network is used to automatically generate wafer defect samples. The method of the present invention can generate a large amount of wafer defect data through small sample data, solving the problem of insufficient deep learning training data caused by the lack of training sets. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0046] Figure 1 A flowchart of a method for generating a wafer defect sample according to the present invention;
[0047] Figure 2 The figure is a schematic diagram of the principle of a method for generating wafer defect samples according to the present invention. DETAILED DESCRIPTION
[0048] The following will be combined with the 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 described embodiments 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 creative work are within the scope of protection of the present invention.
[0049] The purpose of the present invention is to provide a method and system for generating wafer defect samples to solve the problem of insufficient deep learning training data caused by the lack of training sets.
[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0051] Figure 1 is a flow chart of a method for generating wafer defect samples according to the present invention, Figure 2 The schematic diagram of the principle of a method for generating wafer defect samples of the present invention is shown in FIG. Figure 1 and Figure 2 , a method for generating a wafer defect sample of the present invention comprises:
[0052] Step 101: Obtain a sample set of marked wafer defect images.
[0053] Before obtaining the marked wafer defect image sample set in step 101, the wafer defect image samples are manually marked. If a pixel point has a defect, it is marked as 1, otherwise it is marked as 0. Then the marked wafer defect image samples are obtained to form the marked wafer defect image sample set. Therefore, the wafer defect image sample set includes multiple wafer defect images; the pixel point with a defect in each wafer defect image is marked as 1, and the pixel point without a defect is marked as 0.
[0054] Step 102: Divide the wafer defect image sample set into a training set and a test set.
[0055] 50% of all wafer defect images are randomly selected as training sets, and the other 50% are used as test sets. In other words, 50% of the wafer defect images are randomly selected from the wafer defect image sample set as training sets, and the remaining 50% of the wafer defect images are used as test sets.
[0056] Step 103: Establish a radial basis neural network.
[0057] A radial basis neural network is established, whose network input is the pixel points of the wafer defect image, and the network output is the probability of the defect. The radial basis neural network has one hidden layer, including K radial basis, and the connection mode is full connection. The radial basis neural network formula is as follows:
[0058]
[0059]
[0060]
[0061] Where x is the input of the radial basis neural network, i.e., the pixel point of the wafer defect image. i is the radial basis center, η i is the radial basis width. i (x) is the Gaussian radial basis, w i is the network weight, K is the number of radial basis contained in the hidden layer, f(x) is the output of the hidden layer, and Net(x) is the network output, i.e., the defect probability.
[0062] Step 104: Perform model training on the radial basis function neural network using the training set to generate a trained radial basis function neural network.
[0063] The training set obtained in step 102 is used as the training data of the radial basis neural network, and the radial basis neural network is trained using the gradient descent method to update its network parameters to generate the trained radial basis neural network. Assuming that the radial basis neural network parameter set is θ, the parameter update formula is:
[0064]
[0065] In the formula, θ new , is the updated parameter of the radial basis neural network, θ old is the current parameter of the radial basis neural network, α is the learning rate of the radial basis neural network, Net(x) is the radial basis neural network, is the radial basis neural network parameter gradient.
[0066] Step 105: Utilize the trained radial basis function neural network to calculate the test set to obtain a defect generation training set.
[0067] The radial basis function neural network obtained in step 104 is used to calculate the test set to obtain the probability that each sample in the test set is a defect, and the defects with a probability greater than 50% are classified into the defect generation training set. That is, the defect probability of each sample in the test set is calculated using the trained radial basis function neural network obtained in step 104, and the samples with a defect probability greater than 50% are classified into the defect generation training set, thereby obtaining a new defect generation training set.
[0068] Step 106: Use the defect probability as network input, use the wafer defect image pixel defect as network output, and reconstruct the radial basis neural network.
[0069] A radial basis neural network identical to step 103 is re-established, and the network input and output are swapped, that is, the network input is the defect probability, and the network output is the image pixel defect.
[0070] Step 107: Use the defect generation training set to train the reconstructed radial basis function neural network to generate a trained radial basis function neural network.
[0071] The defect generation training set obtained in step 105 is used as training data for the radial basis function neural network established in step 106, and the training is performed in the same manner as in step 104, and the network parameters are updated to obtain a trained radial basis function neural network.
[0072] Step 108: Automatically generate wafer defect samples using the trained radial basis function neural network.
[0073] Using the radial basis function neural network obtained in step 107, different probabilities (e.g., 0.5-1) are input for calculation to generate defects, and the defects are added to the normal image to automatically generate defect samples. That is, different probabilities are input to the trained radial basis function neural network for calculation to generate corresponding pixel defects, and the pixel defects are added to the normal wafer image to automatically generate the wafer defect samples.
[0074] The wafer defect sample generation method provided by the present invention first divides the wafer defect image into a training set and a test set at a ratio of 50% each; secondly, the radial basis neural network model is trained through the training set, and the defect probability of the test set is calculated using the trained model, and the samples with a defect probability greater than 50% are classified into the defect generation training set; then the radial basis neural network is reconstructed, the network input and output are swapped, and the defect generation training set is used for training; finally, different defects are generated according to the input defect probability, and added to the normal image to realize the automatic generation of wafer defect samples, which can solve the problem of insufficient deep learning training data caused by the lack of training sets. The method of the present invention can automatically generate a sufficient number of defect samples for use by an artificial intelligence algorithm based on deep learning, thereby realizing wafer defect detection.
[0075] In addition, the radial basis neural network used in step 103 of the method of the present invention has the characteristics of strong generalization ability and fast convergence speed compared with the traditional BP neural network. Steps 104 to 105 of the method train the radial basis neural network model through the training set, and then use the trained model to calculate the test set, using continuous numerical values to represent the probability of the existence of defects, which solves the problem that the traditional sample set can only express whether there is a defect. Steps 106 to 108 of the method generate different defects by inputting different defect probabilities and add them to the normal picture, which enhances the diversity of samples.
[0076] Based on the wafer defect sample generating method provided by the present invention, the present invention also provides a wafer defect sample generating system, the system comprising:
[0077] A wafer defect image acquisition module is used to acquire a marked wafer defect image sample set; the wafer defect image sample set includes multiple wafer defect images; a pixel point with a defect in each wafer defect image is marked as 1, and a pixel point without a defect is marked as 0;
[0078] A wafer defect image classification module, used for dividing the wafer defect image sample set into a training set and a test set;
[0079] A radial basis neural network establishment module is used to establish a radial basis neural network; the input of the radial basis neural network is the pixel points of the wafer defect image, and the output of the radial basis neural network is the defect probability;
[0080] A radial basis neural network training module, used for performing model training on the radial basis neural network through the training set to generate a trained radial basis neural network;
[0081] A defect generation training set acquisition module, used to calculate the test set using the trained radial basis neural network to obtain a defect generation training set;
[0082] A radial basis neural network reconstruction module is used to use the defect probability as a network input and the wafer defect image pixel point defect as a network output to reconstruct a radial basis neural network;
[0083] A radial basis neural network training module is reconstructed to train the reconstructed radial basis neural network using the defect generation training set to generate a trained radial basis neural network;
[0084] The wafer defect sample automatic generation module is used to automatically generate wafer defect samples using the trained radial basis function neural network.
[0085] Among them, the wafer defect image classification module specifically includes:
[0086] The wafer defect image classification unit is used to randomly select 50% of the wafer defect images from the wafer defect image sample set as a training set and the remaining 50% of the wafer defect images as a test set.
[0087] The radial basis neural network training module specifically includes:
[0088] The radial basis neural network training unit is used to use the training set as training data of the radial basis neural network, use the gradient descent method to train the radial basis neural network, update the network parameters, and generate the trained radial basis neural network.
[0089] The defect generation training set acquisition module specifically includes:
[0090] A defect probability calculation unit, used to calculate the defect probability of each sample in the test set using the trained radial basis neural network;
[0091] The defect generation training set acquisition unit is used to classify the samples with defect probability greater than 50% into the defect generation training set.
[0092] The wafer defect sample automatic generation module specifically includes:
[0093] Pixel point defect generation correspondence, used to input different probabilities into the trained radial basis neural network for calculation to generate corresponding pixel point defects;
[0094] The wafer defect sample generating unit is used to add the pixel point defect into the normal wafer image to generate the wafer defect sample.
[0095] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0096] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A method for generating wafer defect samples, characterized in that: include: Obtain a sample set of marked wafer defect images; The wafer defect image sample set includes a plurality of wafer defect images; In each of the wafer defect images, pixels with defects are marked as 1, and pixels without defects are marked as 0; Dividing the wafer defect image sample set into a training set and a test set; Establishing a radial basis function neural network; the input of the radial basis function neural network is the pixel points of the wafer defect image, and the output of the radial basis function neural network is the defect probability; Performing model training on the radial basis neural network using the training set to generate a trained radial basis neural network; Utilizing the trained radial basis neural network to calculate the test set, to obtain a defect generation training set; Using the defect probability as a network input and the pixel defect of the wafer defect image as a network output, reconstructing the radial basis neural network; Using the defect generation training set to train the reconstructed radial basis neural network to generate a trained radial basis neural network; The trained radial basis function neural network is used to automatically generate wafer defect samples.
2. The method according to claim 1, characterized in that The step of dividing the wafer defect image sample set into a training set and a test set specifically includes: 50% of the wafer defect images are randomly selected from the wafer defect image sample set as a training set, and the remaining 50% of the wafer defect images are selected as a test set.
3. The method according to claim 1, characterized in that The performing model training on the radial basis neural network by using the training set to generate a trained radial basis neural network specifically includes: The training set is used as training data of the radial basis neural network, the radial basis neural network is trained using a gradient descent method, network parameters are updated, and the trained radial basis neural network is generated.
4. The method according to claim 1, characterized in that The method of calculating the test set by using the trained radial basis function neural network to obtain a defect generation training set specifically includes: Calculating the defect probability of each sample in the test set using the trained radial basis neural network; The samples with defect probability greater than 50% are classified into the defect generation training set.
5. The method according to claim 1, characterized in that The method of automatically generating wafer defect samples by using the trained radial basis function neural network specifically includes: Inputting different probabilities into the trained radial basis neural network for calculation to generate corresponding pixel defects; The pixel defect is added to a normal wafer image to generate the wafer defect sample.
6. A wafer defect sample generation system, characterized in that: include: A wafer defect image acquisition module is used to acquire a sample set of marked wafer defect images; The wafer defect image sample set includes a plurality of wafer defect images; In each of the wafer defect images, pixels with defects are marked as 1, and pixels without defects are marked as 0; A wafer defect image classification module, used for dividing the wafer defect image sample set into a training set and a test set; A radial basis neural network establishment module is used to establish a radial basis neural network; the input of the radial basis neural network is the pixel points of the wafer defect image, and the output of the radial basis neural network is the defect probability; A radial basis neural network training module, used for performing model training on the radial basis neural network through the training set to generate a trained radial basis neural network; A defect generation training set acquisition module, used to calculate the test set using the trained radial basis neural network to obtain a defect generation training set; A radial basis neural network reconstruction module is used to use the defect probability as a network input and the wafer defect image pixel point defect as a network output to reconstruct a radial basis neural network; A radial basis neural network training module is reconstructed to train the reconstructed radial basis neural network using the defect generation training set to generate a trained radial basis neural network; The wafer defect sample automatic generation module is used to automatically generate wafer defect samples using the trained radial basis function neural network.
7. The system according to claim 6, characterized in that The wafer defect image classification module specifically includes: The wafer defect image classification unit is used to randomly select 50% of the wafer defect images from the wafer defect image sample set as a training set and the remaining 50% of the wafer defect images as a test set.
8. The system according to claim 6, characterized in that The radial basis neural network training module specifically includes: The radial basis neural network training unit is used to use the training set as training data of the radial basis neural network, use the gradient descent method to train the radial basis neural network, update the network parameters, and generate the trained radial basis neural network.
9. The system according to claim 6, characterized in that The defect generation training set acquisition module specifically includes: A defect probability calculation unit, used to calculate the defect probability of each sample in the test set using the trained radial basis neural network; The defect generation training set acquisition unit is used to classify the samples with defect probability greater than 50% into the defect generation training set.
10. The system according to claim 6, characterized in that The wafer defect sample automatic generation module specifically includes: Pixel point defect generation correspondence, used to input different probabilities into the trained radial basis neural network for calculation to generate corresponding pixel point defects; The wafer defect sample generating unit is used to add the pixel point defect into the normal wafer image to generate the wafer defect sample.