CN117422967A - Wafer map defect mode identification method based on information entropy self-adaptive decision fusion - Google Patents

Wafer map defect mode identification method based on information entropy self-adaptive decision fusion Download PDF

Info

Publication number
CN117422967A
CN117422967A CN202311453152.2A CN202311453152A CN117422967A CN 117422967 A CN117422967 A CN 117422967A CN 202311453152 A CN202311453152 A CN 202311453152A CN 117422967 A CN117422967 A CN 117422967A
Authority
CN
China
Prior art keywords
fusion
wafer map
information entropy
convolution
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202311453152.2A
Other languages
Chinese (zh)
Inventor
陈寿宏
陆颖
覃冠翔
侯杏娜
刘美旗
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guilin University of Electronic Technology
Original Assignee
Guilin University of Electronic Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guilin University of Electronic Technology filed Critical Guilin University of Electronic Technology
Priority to CN202311453152.2A priority Critical patent/CN117422967A/en
Publication of CN117422967A publication Critical patent/CN117422967A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Multimedia (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Testing Or Measuring Of Semiconductors Or The Like (AREA)

Abstract

The invention provides a wafer map defect mode identification method based on information entropy self-adaptive decision fusion, which combines an information entropy self-adaptive decision fusion deep learning model with the characteristic advantage of different Deep Convolutional Neural Networks (DCNNs) to be used for wafer map defect mode identification. Different DCNNs have advantages on different defect mode recognition effects, reasonable weights are given to feature self-adaption extracted by full-connection layers of different DCNs through calculation of information entropy, and then decision fusion is carried out on output results to obtain final fusion features. The fusion characteristics are more expressive, and the final wafer map identification and classification accuracy can be further improved. And finally, inputting the fusion characteristics into an error correction output code and support vector machine (ECOC-SVM) combined model to realize wafer map defect pattern recognition and classification. The invention can accurately identify and analyze the defect mode of the wafer map, is beneficial to improving the yield of wafer production, and has great practical significance for the development of national semiconductor manufacturing technology.

Description

Wafer map defect mode identification method based on information entropy self-adaptive decision fusion
Technical Field
The invention relates to the field of wafer defect detection in the semiconductor manufacturing process, in particular to research on a wafer map defect pattern recognition classification algorithm based on feature fusion.
Background
Semiconductors are an important basis for the electronic information industry, and as semiconductors develop, the complexity of the manufacturing process increases, and problems in the semiconductor manufacturing process cause failures. Wafer fabrication is an important step in semiconductor fabrication, in which wafer defects are generated by the fabrication process, and the wafer defect ratio affects the wafer yield. Wafer testing is a link in wafer fabrication, and integrated circuits are tested according to electrical parameters, and by means of circuit probe testing, the quality of each chip on the wafer is evaluated. Marking the die that passed the functional test, the defective die, and the locations where there are no dies with different colors can form a pattern of a shape pattern, i.e., a wafer map, that reflects potential anomalies in the wafer fabrication process. Through identifying the defect mode of the wafer map, a fault source existing in the semiconductor manufacturing process is positioned, engineers are helped to find abnormal operation in time and adjust the abnormal operation, the production efficiency and the production quality of the wafer are improved, and loss caused by occurrence of a large number of defects is avoided.
In the conventional wafer map inspection method, an experienced semiconductor engineer identifies a defect mode by analyzing information such as the size, shape, position and the like of a wafer map defect through visual identification, so as to analyze the actual defect cause. But this method is not only costly, time-consuming and labor-consuming, but also of low accuracy. With the progress of hardware and algorithm, the problem of high cost, long time and low accuracy of manual detection can be overcome by applying the shallow machine learning method to wafer map defect detection. But due to the construction of focusing on the defect pattern identifier, no effective features can be learned from a complex and diverse image. With the development of artificial intelligence, a method of deep learning becomes one of the mainstream methods of image processing. CNN is widely used in the field of wafer map defect mode classification. The CNN can automatically learn rich features, avoid the problems of incomplete manual feature extraction and significant feature loss of certain categories, and can realize accurate identification and classification of the features.
Disclosure of Invention
The invention aims to provide a wafer map defect mode identification method based on information entropy self-adaptive decision fusion, which combines the identification advantages of different DCNN models on different wafer defect modes, and improves the accuracy of identifying the wafer map defect modes.
In order to achieve the above object, the present invention provides a wafer map defect pattern recognition method, including:
step 1, carrying out data expansion on a wafer map data set image, and dividing the wafer map data set image into a training set and a testing set;
step 2, building four DCNN models for extracting wafer map defect characteristics;
step 3, adopting an information entropy self-adaptive decision fusion method for the features extracted by the four model full-connection layers, and carrying out self-adaptive fusion on the four features by calculating the information entropy and fusion weight to obtain fusion feature information;
and step 4, inputting the fusion characteristics into an ECOC-SVM classifier to carry out final classification, so as to realize the identification and classification of the wafer map defect modes.
In step 1, the number of the four types of wafer maps with a small number is respectively extended to 2500 in the data extension process of the wafer map. The wafer patterns with different sizes are uniformly adjusted to 224,224 and 3, and the channel number is 3; the data set is divided into training and test sets at a ratio of 8:2.
In the step 2, four built DCNN models are respectively three DCNN models based on a characteristic pyramid (FPN), and one DCNN model based on a multi-branch convolution block attention module (M-CBAM);
the first DCNN model based on the feature pyramid has a total convolution layer number of 10 layers; extracting three layers of features in the 6 th convolution layer, the 8 th convolution layer and the 10 th convolution layer respectively to form a feature pyramid; the deep features are up-sampled and then fused with the shallow features to obtain three new fusion features, and finally the shallow features and the three new fusion features are further spliced and fused;
the second DCNN model based on the feature pyramid has the total convolution layer number of 27 layers, and the first layer of convolution receives the input of the wafer map data set; the second layer convolution kernel size is 3×3; the back is 4 residual structures, each residual structure is provided with two residual blocks, and each residual block is composed of three convolution layers; extracting three features from the output layers of the second residual structure, the third residual structure and the fourth residual structure to form a feature pyramid; the deep features are up-sampled and then fused with the shallow features to obtain three new fusion features, and finally the shallow features and the three new fusion features are further spliced and fused;
the third DCNN model based on the feature pyramid adopts ResNet50 with deeper layers as a feature extraction network, the feature extraction at the output layer of each residual block forms the feature pyramid, the deep features are up-sampled and then fused with the shallow features to obtain four new fusion features, and finally the shallow features and the four new fusion features are further spliced and fused;
the fourth DCNN model based on the multi-branch convolution block attention module adopts ResNeXt50 as a feature extraction network, and fuses M-CBAM into each residual block of the ResNeXt 50. Each ResNeXt block is fused into M-CBAM, and richer features are extracted.
In step 3, features are extracted at four DCNN full connection layers, and one-dimensional vectors are converted into probability vectors through a Softmax layer, wherein each probability corresponds to one category. The DCNN information entropy self-adaptive decision fusion method comprises the steps of firstly calculating four DCNN output probability matrixes, then calculating the information entropy of each DCNN output probability vector, and giving reasonable weight according to the information entropy; further, multiplying the output probability of each DCNN with the fusion weight to obtain a new probability matrix, and finally, carrying out column weighted summation on the new probability matrix to obtain the final fusion feature.
In step 4, the ECOC-SVM combines the error correction output code with the support vector machine, and simultaneously adopts a plurality of SVMs to convert the multi-classification problem into a plurality of two-classification problems, thereby realizing multi-classification and having an error correction function. The fusion characteristics obtained after the four DCNN information entropy self-adaptive decision fusion are input to the ECOC-SVM, so that the final wafer map recognition and classification accuracy can be further improved.
Drawings
FIG. 1 is a schematic diagram of a method for identifying a defect mode of a wafer map based on information entropy adaptive decision fusion;
FIG. 2 is a flow chart of a method for identifying a defect mode of a wafer map based on information entropy adaptive decision fusion;
FIG. 3 is a first feature pyramid based FPN-DCNN11 model in accordance with the present invention;
FIG. 4 is a second feature pyramid based FPN-DCNN-Res27 model in accordance with the present invention;
FIG. 5 is a third feature pyramid based FPN-Resnet50 model in accordance with the present invention;
FIG. 6 is a fourth M-CBAM-ResneXt50 model based on a multi-branch convolution block attention module according to the present invention;
FIG. 7 is a multi-branch convolution block attention module according to the present disclosure;
Detailed Description
Embodiments of the present invention are described below by way of specific examples.
Referring to fig. 1, a schematic diagram of a method for identifying a defect mode of a wafer map based on information entropy adaptive decision fusion is provided.
As shown in fig. 2, specifically, the method for identifying the defect mode of the information entropy-based adaptive decision fusion wafer map includes the following steps:
step S201: the four wafer patterns of Donut, near-full, random and Scratch are respectively expanded to 2500 wafers in a rotating mode, the wafer patterns with different sizes are uniformly adjusted to be (224, 3), and the channel number is 3. Dividing the data set into a training set and a testing set according to the proportion of 8:2;
step S202: respectively building four DCNN models based on a feature pyramid and multi-branch attention convolution as a feature extraction backbone network;
step S203: inputting a training data set into four DCNN models for learning training to obtain extracted characteristic information, and then adopting DCNN information entropy self-adaptive decision fusion to obtain fusion characteristics;
step S204: and inputting the fusion characteristics into an ECOC-SVM classifier to carry out final classification, so as to realize the identification and classification of the wafer map defect modes.
As shown in fig. 3, the building of the first DCNN model based on the feature pyramid specifically includes:
FPN-DCNN11 has a total of 10 convolution layers, the first convolution receives an input of a wafer map dataset, and the first convolution comprises a convolution layer, a batch normalization layer and a ReLU activation function; the dimension of the wafer map of the input convolution layer is [3,224,224], and the size of the convolution kernel is 7×7; the remaining nine layers of convolutions use a convolution kernel size of 3 x 3; in order to fuse the shallow and deep feature information, three layers of features C6, C8, and C10 are extracted from the 6 th, 8 th, and 10 th convolution layers of DCNN11, respectively, to form a feature pyramid. And fusing the deep features P10 with the shallow features C8 to obtain new features P8 containing deep semantic feature information. Further, the P8 and the C6 are subjected to feature fusion to obtain P6, and finally, the four features of the P6, the P8, the P10 and the C10 are respectively subjected to global average pooling layer and full connection layer to obtain fusion features by further splicing, and finally, classification is realized through the full connection layer;
as shown in fig. 4, the building of the second DCNN model based on the feature pyramid specifically includes:
the FPN-DCNN-Res27 deepens the number of network layers by the residual structure, and retains two convolutional layers in front of the FPN-DCNN11, and 4 residual structures behind, each residual structure has two residual blocks, each residual block has a total number of 27 layers by using convolutional layers with convolution kernels of 1×1,3×3, and 1×1. Three features C4, C5 and C6 are extracted from the output layers of the second residual structure, the third residual structure and the fourth residual structure of the FPN-DCNN-Res27, a feature pyramid is constructed, the deep features are up-sampled and then fused with the shallow features, and finally fusion features P4, P5 and P6 are obtained. The four features P4, P5, P6 and C5 are respectively subjected to feature splicing after passing through a global average pooling layer and a full connection layer, and finally are classified through the full connection layer;
as shown in fig. 5, the building of the third DCNN model based on the feature pyramid specifically includes:
the FPN-ResNet50 adopts a network model ResNet50 with a deeper layer number as a reference network; the Resnet is divided into five phases, the first phase being a convolution layer with a convolution kernel of size 7 x 7 for the preliminary extraction of features. The four later stages are residual structures, each comprising a number of residual blocks [3,4,6,2], and deep features are further extracted through a deeper residual network. The second stage reduces the feature map size by half through a pooling layer with a pooling core of 3 x 3. In the following three stages, the first residual block of each stage is downsampled, so that the size of the feature map is reduced by half. Extracting features at each residual block output layer of the DCNN-ResNet50, constructing a feature pyramid, upsampling deep features, fusing the deep features with shallow features, and finally obtaining C5, P5, P4, P3 and P2. The five features are subjected to feature splicing after passing through a global average pooling layer and a full connection layer, and finally classification is realized through the full connection layer;
as shown in fig. 6, the building a DCNN model based on a fourth multi-branch convolution block attention module specifically includes:
M-CBAM-ResneXt50 adopts ResNeXt50 as a feature extraction network, and M-CBAM is fused into each residual block of ResNeXt50 in order to pertinently improve the feature extraction capability of different types of defect wafer maps. Each ResNeXt block is integrated with M-CBAM, so that different weights are given to different types of wafer defect clusters when characteristic information is extracted by the convolutional neural network, and the convolutional neural network focuses on main defect cluster positions from channels and spaces to restrain the influence of noise and extract richer characteristics. After a series of convolution blocks, the network extracts deep abstract and high-level semantic features, and finally final features are obtained after a global average pooling layer and a full connection layer.
Specifically, as shown in fig. 7, the multi-branch convolution block attention module is:
the multi-branch convolution block attention module includes a multi-branch channel attention module and a dual-branch spatial attention module. The multi-branch channel attention module is provided with three branches, wherein the first branch is the maximum pooling operation, the second branch is the average pooling operation, and the third branch is the characteristic obtained by the maximum pooling and the average pooling for splicing. The characteristics obtained by the three branches are compressed in the channel dimension through the multi-layer perceptron to obtain the characteristic information of the important channels, and then the characteristic information is subjected to dimension increasing operation to obtain the number of channels with dimension of C, so that the output characteristic information containing the important channels is obtained. The multi-branch channel attention module expression is:
wherein F inputs a feature map, M C (F) In the case of a multi-leg channel attention module,representing the multiplication of the feature map elements, F' being the output feature map. />Representing the splice characteristics after the maximum pooling and average pooling operations. W (W) 0 ∈R C×C/r ，W 0 ′∈R 2C ×C/r And W is 1 ∈R C/r×C Representing the weights of a Shared multi-layer perceptron (Shared MLP), σ is the Sigmoid activation function. The characteristics with larger channel weight are screened out through the multi-branch channel attention module, and more accurate characteristic information is obtained.
The double-branch space attention module is provided with two branches, two features are obtained by respectively carrying out maximum pooling and average pooling on the input feature map, the two features are spliced, and then a convolution operation is carried out by using a convolution kernel of 7 multiplied by 7 to extract the features. The expression of the double-branch spatial attention module is as follows:
wherein F' is the output characteristic of the attention module through the multi-branch channel, M S (F) Is a dual-branch spatial attention module,representing the multiplication of the feature map elements, F "being the final output feature map. />And->The operation outputs of average pooling and maximum pooling, f 7×7 For a convolution kernel of 7 x 7, σ is the Sigmoid activation function. And through the double-branch space attention module, the obvious positions of the defect clusters are classified with larger weights in space, so that the characteristic information which is concentrated on the space positions of the main defect clusters is obtained.
The DCNN information entropy self-adaptive decision fusion algorithm specifically comprises the following steps:
step 1: four DCNN output probability matrices are calculated. The wafer map is respectively input into four DCNNs, the features extracted from the full connection of each DCNN are converted into probability vectors through a Softmax function, and the four probability vectors form a probability matrix:
wherein x is an input wafer map, P (x) is a probability output value, and n is a category number. Each row in the matrix represents the probability value output by each DCNN corresponding category, and the category with the largest probability value is the category predicted by the DCNN;
step 2: and calculating the information entropy of the four DCNN output probability matrixes. The uncertainty of the DCNN on the input image prediction can be represented by information entropy, and the information entropy of each DCNN output probability vector is calculated by the following specific formula:
wherein H is i (x) Information entropy, P representing i-th DCNN output probability ij (x) Representing the predicted probability of the ith DCNN for the jth class. The larger the information entropy is, the larger the representative uncertainty is, namely, the uncertainty of the DCNN on the prediction classification result of the input image is.
Step 3: and calculating the self-adaptive fusion weights of the four DCNN information entropies. The larger the information entropy is, the more uncertain the classification effect is, and the smaller the weight is given to the DCNN. The self-adaptive fusion weight calculation formula is as follows:
wherein H is i (x) Information entropy, W representing i-th DCNN output probability i The fusion weight representing the i-th DCNN output probability.
Step 4: and (5) weighting and fusing by a decision layer. After the fusion weight of each DCNN output probability is calculated, multiplying the output probability of each DCNN by the fusion weight to obtain a new probability matrix:
the new probability matrix is added with self-adaptive fusion weights, and DCNNs with reliable comparison and better classification effect are endowed with larger weights through different weights endowed to different DCNN output probabilities. The new probability matrix is weighted and summed according to the columns to obtain the characteristics after fusion:
P″(x)＝[P 1 (x)P 2 (x)P 3 (x)···P n (x)]
through DCNN information entropy self-adaptive decision fusion, reasonable fusion weights can be given to different models, so that the optimal classification effect is decided.

Claims (6)

1. The wafer map defect mode identification method based on information entropy self-adaptive decision fusion is characterized by comprising the following steps of:
step 1, carrying out data expansion on a wafer map data set image, and dividing the wafer map data set image into a training set and a testing set;
step 2, constructing four deep convolution neural networks based on feature pyramids and multi-branch convolution block attention modules as wafer map defect feature extraction backbone networks;
extracting features at all connection layers of the four models respectively, adopting an information entropy self-adaptive decision fusion method, and carrying out self-adaptive fusion on the four features by calculating information entropy and fusion weight to obtain fusion feature information;
and step 4, inputting the fusion characteristics into an ECOC-SVM classifier to carry out final classification, so as to realize the identification and classification of the wafer map defect modes.
2. The method for identifying the defect mode of the wafer map based on the information entropy self-adaptive decision fusion according to claim 1, wherein the wafer maps with different sizes are uniformly adjusted to 224×224, and the pixel channel is an image with 3; the ratio of the whole data set to the training set to the test set is 8:2.
3. The method for identifying the defect mode of the wafer map based on the information entropy self-adaptive decision fusion according to claim 1, wherein four deep convolution neural networks based on feature pyramids and multi-branch convolution block attention modules are built as a wafer map defect feature extraction backbone network, and the method comprises the following steps:
3-1, the first depth convolution neural network based on the feature pyramid is characterized in that the total convolution layer number is 10; extracting three layers of features in the 6 th convolution layer, the 8 th convolution layer and the 10 th convolution layer respectively to form a feature pyramid;
3-2, the second depth convolution neural network based on the feature pyramid is characterized in that the total convolution layer number is 27; the first 2 layers of convolution layers of the first depth convolution neural network are reserved, 4 residual structures are arranged behind the first 2 layers of convolution layers, each residual structure has 2 residual blocks, and each residual block comprises 3 convolution layers; outputting the extracted three-layer features in the second, third and fourth residual structures to form a feature pyramid;
3-3, the third depth convolution neural network based on the feature pyramid is characterized in that the total convolution layer number is 50; adopting ResNet50 with deeper layers as a reference network, wherein the first stage is 1 convolution layer, and the later four stages are residual structures; four layers of features extracted by each residual structure output layer form a feature pyramid; 3-4, the fourth depth convolution neural network based on the multi-branch convolution block attention module is characterized in that the total convolution layer number is 50; resNeXt50 is used as a reference network and a multi-branch convolution block attention module is incorporated into each residual block of ResNeXt 50.
4. The method of claim 3, wherein the dimension of the wafer map is [3,224,224], 3 represents the number of channels of the feature vector, 224 represents the height of the feature vector, and 224 represents the width of the feature vector.
5. The method for identifying a defect mode of a wafer map based on information entropy adaptive decision fusion according to claim 1, wherein the method for adopting the information entropy adaptive decision fusion comprises the following steps:
5-1. Calculating the four DCNN output probability matrices of claim 3;
5-2. Calculating the information entropy of the four DCNNs according to claim 3;
5-3. According to the information entropy of the four DCNNs of claim 3, the fusion weights are adaptively used;
5-4, weighting and fusing by a decision layer; after the fusion weight of each DCNN output probability is calculated, multiplying the output probability of each DCNN by the fusion weight to obtain a new probability matrix; and the new probability matrix is weighted and summed according to the columns to obtain the fused characteristic.
6. The wafer map defect mode identification method based on information entropy adaptive decision fusion according to claim 1, wherein Error Correction Output Coding (ECOC) and Support Vector Machines (SVM) are combined, and a plurality of SVMs are adopted to convert a multi-classification problem into a plurality of two-classification problems, so that multi-classification is realized and an error correction function is provided; the fusion features obtained after the four kinds of DCNN information entropy adaptive decision fusion in claim 3 are input to an ECOC-SVM.
CN202311453152.2A 2023-11-03 2023-11-03 Wafer map defect mode identification method based on information entropy self-adaptive decision fusion Pending CN117422967A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311453152.2A CN117422967A (en) 2023-11-03 2023-11-03 Wafer map defect mode identification method based on information entropy self-adaptive decision fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311453152.2A CN117422967A (en) 2023-11-03 2023-11-03 Wafer map defect mode identification method based on information entropy self-adaptive decision fusion

Publications (1)

Publication Number Publication Date
CN117422967A true CN117422967A (en) 2024-01-19

Family

ID=89528142

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311453152.2A Pending CN117422967A (en) 2023-11-03 2023-11-03 Wafer map defect mode identification method based on information entropy self-adaptive decision fusion

Country Status (1)

Country Link
CN (1) CN117422967A (en)

Similar Documents

Publication Publication Date Title
CN109873610B (en) Photovoltaic array fault diagnosis method based on IV characteristic and depth residual error network
CN110766660A (en) Integrated circuit defect image recognition and classification system based on fusion depth learning model
CN112561910A (en) Industrial surface defect detection method based on multi-scale feature fusion
CN113642445B (en) Hyperspectral image classification method based on full convolution neural network
CN115375635A (en) Wafer defect identification method
CN113763364B (en) Image defect detection method based on convolutional neural network
CN113781466A (en) Solar panel defect identification method based on multi-scale attention-intensive network
CN117422967A (en) Wafer map defect mode identification method based on information entropy self-adaptive decision fusion
Liu et al. Focal auxiliary classifier generative adversarial network for defective wafer pattern recognition with imbalanced data
CN116486176A (en) Wafer map fault mode identification method based on attention space pyramid pooling
CN115795366A (en) Wafer map fault mode identification method based on multi-branch attention mechanism
CN114494284B (en) Scene analysis model and method based on explicit supervision area relation
CN114155246B (en) Deformable convolution-based power transmission tower pin defect detection method
Song et al. Based on end-to-end object detection algorithm with transformers for detecting wafer maps
CN115641437A (en) Industrial defect instance segmentation method
CN114980723A (en) Fault prediction method and system for cross-working-condition chip mounter suction nozzle
CN114596273A (en) Intelligent detection method for multiple defects of ceramic substrate by using YOLOV4 network
CN114897858A (en) Rapid insulator defect detection method and system based on deep learning
CN113536998A (en) Motor fault diagnosis method and system based on improved stack type sparse self-encoder
CN112634216B (en) Insulator self-explosion detection method based on deep learning model
US20230022057A1 (en) Method for retrieving images from database
CN116740002A (en) Wafer defect detection method based on neural network architecture search
CN117710755B (en) Vehicle attribute identification system and method based on deep learning
TWI834916B (en) Machine learning-based defect detection of a specimen
Huang et al. Imbalanced wafer map dataset classification with semi-supervised learning method and optimized loss function

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination