Integrated circuit shmoo graph classification method based on lightweight deep learning network

Abstract

The invention relates to an integrated circuit shmoo graph classification method based on a lightweight deep learning network, which comprises the following steps of: firstly, acquiring shmoo data of a chip by utilizing automatic test equipment, preprocessing the shmoo data into an image format, then defining classification categories according to different characteristics of the shmoo graph, constructing a data set, and classifying the shmoo graph according to the data set; and finally, classifying the shmoo image by using the lightweight deep learning network designed by the invention. The network architecture designed by the invention has the advantages of smaller network scale, difficulty in overfitting, high training speed and higher overall classification accuracy, and can quickly and accurately realize automatic classification of the shmoo graph.

Description

Technical Field

[0001] The present invention belongs to the field of integrated circuit testing, relates to the field of deep learning, and specifically relates to an integrated circuit shmoo graph classification method based on a lightweight deep learning network. Background Art

[0002] With the rapid development of the semiconductor industry, the complexity of device design and process is increasing, and the cost of the testing link continues to rise in the entire industry. The performance test and quality control of semiconductor devices are key links to ensure product reliability and improve market competitiveness. Only devices that pass the test can enter the mass production stage. As a visualization tool widely used in integrated circuit performance evaluation, the shmoo diagram displays the device status under different test conditions in an intuitive way, so as to determine whether the device works normally under the set conditions. Engineers can set test parameters according to their needs to generate one-dimensional, two-dimensional or even three-dimensional shmoo diagrams. In the common two-dimensional shmoo diagram, the columns correspond to the frequency conditions and the rows correspond to the voltage conditions. Each cell in the diagram records a binary state (pass: 0 or fail: 1), indicating whether the device can work normally under the current voltage and frequency conditions. The shmoo diagram provides engineers with rich information, reveals device characteristics and potential defects, and is of great significance for optimizing test efficiency and improving yield.

[0003] For the same type of semiconductor devices, similar shmoo plots often represent similar device characteristics, process quality, and defect types, so the different categories determined by different shmoo plot features are more important than the test results of a single shmoo unit. As chips become increasingly complex, more categories will appear in shmoo plots. Therefore, it is crucial for engineers to classify shmoo plots and extract useful information from them. Traditional shmoo plot analysis methods rely on engineers' manual identification and classification, which is not only time-consuming and labor-intensive, but also easily affected by subjective judgment, limiting the application of shmoo plots. In recent years, convolutional neural networks have played an important role in classifying images, and deep learning models have become a powerful tool for solving such problems. This method can not only realize the automatic classification of shmoo plots and improve the accuracy of classification, but also significantly improve the ability to process large-scale data sets, which is of great significance for reducing the cost of integrated circuit testing and improving testing efficiency. Summary of the invention

[0004] The purpose of the present invention is to provide an integrated circuit shmoo graph classification method based on a lightweight deep learning network. After collecting the shmoo data of the chip using ATE and preprocessing it into an image format, the shmoo graph is processed and analyzed by deep learning, so that the shmoo graph is automatically classified according to the shmoo features, so as to reduce the workload of test engineers and improve the efficiency of integrated circuit testing.

[0005] To achieve the above object, the technical solution of the present invention is: a shmoo graph classification method for integrated circuits based on a lightweight deep learning network, comprising the following steps:

[0006] Step 1: Use ATE to collect the chip's shmoo raw data;

[0007] Step 2: preprocessing the shmoo raw data to obtain a shmoo graph in image format, so as to facilitate input into a deep learning network for training;

[0008] Step 3: According to the graphic features, the shmoo graph samples obtained in step 2 are classified and labeled to obtain a data set containing 16 shmoo graph categories, and the shmoo data set is divided into a training set and a test set;

[0009] Step 4: Build a deep learning framework based on a lightweight convolutional neural network, which consists of an input layer, five convolutional blocks containing convolutional layers, BN layers and ReLU layers, two maximum pooling layers, one average pooling layer, one spatial pyramid pooling layer, two dropout layers, three fully connected layers and a softmax output layer;

[0010] Step 5: Send the training set images to the deep learning framework for training. After each round of training, send the test set images to record the classification results of the network for the test set and compare them with the real labels of the test set samples, and repeat this process for multiple rounds.

[0011] Step 6: Calculate the evaluation index and compare the classification effect of the lightweight deep learning network with other commonly used deep learning networks.

[0012] The specific implementation process of step 1 is as follows:

[0013] (a) Develop a shmoo test plan based on the chip to be tested. The test plan usually describes how to change the voltage / frequency test conditions at different temperatures;

[0014] (b) According to the test plan, use the ATE software development environment to write the shmoo test program, which involves editing basic forms such as PinMap, ChannelMap, TestFlow, and creating test functions;

[0015] (c) The chip to be tested is placed in a socket on a load board, and the ATE user host loads and executes the shmoo test program, while recording the operating status of the chip under each test condition and storing it in a datalog file in a data format.

[0016] The specific implementation process of step 2 is as follows:

[0017] (a) using a data analysis tool to read the datalog file and obtain the shmoo raw data in .csv format with a dimension of 11x11;

[0018] (b) Use Python to program a data preprocessor to convert the shmoo data in .csv format into .png image format according to the Pass / Fail results in each state. The shmoo cells of Pass are green, and the shmoo cells of Fail are red.

[0019] The specific implementation process of step 3 is as follows:

[0020] (a) The 16 shmoo diagram categories are named "Fail-Normal", "Pass-Normal", "Fail-Central", "Pass-Central", "Fail-Hole", "Fail-Central-Hole", "Fail-X-Wall", "Fail-Y-Wall", "Fail-XY-Wall", " Fail-X-Wall-Central", "Fail-Y-Wall-Central", "Fail-X-Wall-Hole", "Fail-Y-Wall-Hole", "Fail-XY-Wall-Hole", "Fail-X-Wall-Central-Hole", "Fail-Y-Wall-Central-Hole";

[0021] (b) The classification category is composed of four basic features combined with whether it belongs to Pass or Fail. The description of the four basic features is as follows: "Normal" feature refers to a typical Pass / Fail situation without other features, and one area in the shmoo diagram is significantly larger than the other; "Central" feature refers to the Pass / Fail boundary passing through the central area of ​​the shmoo diagram, and the classification is based on which side has a larger proportion; "Hole" feature refers to abnormal scattered points or irregular protrusions / vacancies in the Shmoo diagram, and this feature belongs to Fail; "Wall" refers to a continuous long Fail segment under a certain test condition in the X, Y or XY direction to form a wall-like pattern, and this feature belongs to Fail;

[0022] (c) The shmoo dataset is divided into a training set and a test set, and the ratio of the training set to the test set is 70% and 30% of the dataset respectively.

[0023] The specific implementation process of step 4 is as follows:

[0024] (a) Add an input layer with an input size of 96x96 and 3 channels;

[0025] (b) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 128 channels, and a stride of 1;

[0026] (c) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 64 channels, and a stride of 1;

[0027] (d) Add a maximum pooling layer with a convolution kernel size of 2x2, 64 channels, and a stride of 2;

[0028] (e) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 32 channels, and a stride of 1;

[0029] (f) Add an average pooling layer with a convolution kernel size of 2x2, 32 channels, and a stride of 2;

[0030] (g) Add a dropout layer with a dropout ratio of 0.2;

[0031] (h) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 16 channels, and a stride of 1;

[0032] (i) Add a maximum pooling layer with a convolution kernel size of 2x2, 16 channels, and a stride of 2;

[0033] (j) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 8 channels, and a stride of 1;

[0034] (k) Add a spatial pyramid pooling layer, use multi-scale pooling with adaptive kernel size, and output a 1x168 feature vector;

[0035] (l) Add an 84-bit fully connected layer, a BN layer and a ReLU layer;

[0036] (m) Add a dropout layer with a dropout ratio of 0.2;

[0037] (n) Add a 42-bit fully connected layer, a BN layer and a ReLU layer;

[0038] (o) Add a 16-bit fully connected layer;

[0039] (p) Add an activation layer using the Softmax activation function as the output layer.

[0040] The specific implementation process of step 5 is as follows:

[0041] (a) After the training set is input into the deep learning network, forward propagation is performed, the classification results are output and compared with the true labels of the training set samples, and then the obtained errors are back-propagated to update and correct the parameters of each layer in the network, thus completing a round of training.

[0042] (b) After each round of training, the test set images are fed in, and the classification effect of the network on the test set is recorded and compared with the true labels of the test set samples. Multiple rounds are iterated, and the training accuracy, training loss, test accuracy and test loss in the entire training and testing process are recorded and plotted.

[0043] The specific implementation process of step 6 is as follows:

[0044] (a) Plot a test set classification accuracy comparison curve of the lightweight deep learning network and the commonly used deep learning network running on the shmoo dataset.

[0045] (b) Calculate common evaluation indicators, including macro-average precision, macro-average recall, macro-average F1 score, Matthews correlation coefficient, network parameter number and model size as a comparison supplement.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] In the method of the present invention, the original shmoo data is preprocessed into an image format by using Python programming, and the shmoo classification problem is processed using a deep learning method, which is a good improvement over the traditional classification method;

[0048] The method of the present invention provides a classification category and label name for dividing shmoo graphs according to shmoo features, explains the basic features of shmoo as the basis for division, and can be used as a classification reference in the industry;

[0049] In the method of the present invention, a new lightweight convolutional neural network is used to realize the automatic intelligent classification of shmoo graphs, with fast training speed and high classification accuracy, thus improving the efficiency of integrated circuit testing;

[0050] The method of the present invention uses a smaller network scale and fewer learnable parameters to overcome overfitting, and introduces spatial pyramid pooling, which can not only capture local details and global information in the image, but also help the network learn more generalized feature representations, thereby achieving an ideal shmoo graph classification effect;

[0051] The method of the present invention can realize online real-time shmoo graph classification detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the steps of the present invention.

[0053] Figure 2 It is a flow chart of using ATE to collect shmoo raw data in an embodiment of the present invention.

[0054] FIG3 is a schematic diagram of examples of the classification categories of the shmoo graphs proposed in the present invention, (a) is a “Fail-Normal” class, (b) is a “Pass-Normal” class, (c) is a “Fail-Central” class, (d) is a “Pass-Central” class, (e) is a “Fail-Hole” class, (f) is a “Fail-Central-Hole” class, (g) is a “Fail-X-Wall” class, (h) is a “Fail-Y-Wall” class, and (i) is a “Fail-XY- The following table lists the “Fail-X-Wall-Hole” class, (j) the “Fail-X-Wall-Central” class, (k) the “Fail-Y-Wall-Central” class, (l) the “Fail-X-Wall-Hole” class, (m) the “Fail-Y-Wall-Hole” class, (n) the “Fail-XY-Wall-Hole” class, (o) the “Fail-X-Wall-Central-Hole” class, and (p) the “Fail-Y-Wall-Central-Hole” class.

[0055] Figure 4 It is a structural diagram of a lightweight deep learning network used by the present invention to realize automatic classification of shmoo graphs.

[0056] Figure 5 It is a curve graph of training loss, training accuracy, test loss and test accuracy obtained by running the lightweight deep learning network on the shmoo graph dataset in an embodiment of the present invention.

[0057] Figure 6 It is a comparison chart of test set accuracy curves obtained by running the lightweight deep learning network and the commonly used deep learning network on the shmoo graph dataset in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention is further described below in conjunction with the accompanying drawings and embodiments. The present invention includes but is not limited to the following embodiments.

[0059] The present invention first uses ATE to collect shmoo raw data, and uses a preprocessing program to convert the shmoo data into an image format to obtain a shmoo graph data set, and then labels the data set according to the shmoo graph classification category proposed in the present invention, and then uses the lightweight deep learning network designed by the present invention to classify the shmoo graph, and draws the accuracy curve and loss curve of the network for the training set and the accuracy curve and loss curve of the network for the test set to show the network effect, and finally uses a variety of evaluation indicators to compare with several commonly used deep learning networks to highlight the advantages of the lightweight deep learning network designed by the present invention.

[0060] like Figure 1 As shown, the shmoo graph classification method based on a lightweight deep learning network provided by the present invention comprises the following steps:

[0061] Step 1: ATE collects shmoo raw data. The specific flow chart is as follows Figure 2 shown.

[0062] First, a shmoo test plan is formulated. In the embodiment, the horizontal axis is used as the voltage condition and the vertical axis is used as the frequency condition to collect shmoo data containing 11x11 cells.

[0063] Then, according to the test plan, use the ATE software development environment to write the shmoo test program, combine the specifications of the batch of chips to define the basic forms such as PinMap, ChannelMap, TestFlow, and create the shmoo test function.

[0064] Finally, the chip to be tested is placed in the socket on the load board, and the user host loads and executes the shmoo test program to record the operating status of the chip under each given test condition, and the results are stored in a datalog file in a data format.

[0065] Step 2: Preprocess the shmoo image into image format.

[0066] The raw shmoo data in .csv format is preprocessed to obtain shmoo graphs in .png format that can be input into the deep learning network.

[0067] In the embodiment, a total of 1078 shmoo data samples were obtained from the batch of test chips, and 1078 shmoo image samples were obtained after preprocessing, constituting the data set of the embodiment.

[0068] Step 3: Define shmoo classification categories and build the dataset.

[0069] According to the four basic characteristics of the shmoo graph and the combination with the Pass / Fail results, the shmoo graph is divided into 16 categories. Example diagrams of each category are shown in Figure 3.

[0070] The four basic features are: "Normal" feature refers to a typical Pass / Fail situation without other features, and one area in the shmoo diagram is significantly larger than the other; "Central" feature refers to the Pass / Fail boundary passing through the central area of ​​the shmoo diagram, and is classified according to which side accounts for a larger proportion; "Hole" feature refers to abnormal scattered points or irregular protrusions / gaps in the Shmoo diagram, etc., all of which belong to Fail; "Wall" refers to a continuous long Fail segment under a certain test condition in the X, Y or XY direction to form a wall-like pattern, all of which belong to Fail.

[0071] In the embodiment, the number of each label classification of the 1078 shmoo graph samples is as follows: 42 for “Fail-Normal”, 223 for “Pass-Normal”, 42 for “Fail-Central”, 46 for “Pass-Central”, 221 for “Fail-Hole”, 49 for “Fail-Central-Hole”, 48 for “Fail-X-Wall”, 52 for “Fail-Y-Wall”, and 53 for “Fail-XY-Wall”. l” category, 31 pictures, 55 pictures of “Fail-X-Wall-Central” category, 50 pictures of “Fail-Y-Wall-Central” category, 55 pictures of “Fail-X-Wall-Hole” category, 55 pictures of “Fail-Y-Wall-Hole” category, 33 pictures of “Fail-XY-Wall-Hole” category, 38 pictures of “Fail-X-Wall-Central-Hole” category, and 38 pictures of “Fail-Y-Wall-Central-Hole” category.

[0072] The data set of the embodiment is divided into a training set and a test set, and the division ratios of the training set and the test set are 70% and 30% of the data set respectively.

[0073] Step 4: Build a lightweight deep learning network. In an embodiment, the lightweight deep learning network structure is as follows: Figure 4 The specific establishment method is as follows:

[0074] (a) Add an input layer with an input size of 96x96 and 3 channels;

[0075] (b) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 128 channels, and a stride of 1;

[0076] (c) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 64 channels, and a stride of 1;

[0077] (d) Add a maximum pooling layer with a convolution kernel size of 2x2, 64 channels, and a stride of 2;

[0078] (e) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 32 channels, and a stride of 1;

[0079] (f) Add an average pooling layer with a convolution kernel size of 2x2, 32 channels, and a stride of 2;

[0080] (g) Add a dropout layer with a dropout ratio of 0.2;

[0081] (h) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 16 channels, and a stride of 1;

[0082] (i) Add a maximum pooling layer with a convolution kernel size of 2x2, 16 channels, and a stride of 2;

[0083] (j) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 8 channels, and a stride of 1;

[0084] (k) Add a spatial pyramid pooling layer, use multi-scale pooling with adaptive kernel size, and output a 1x168 feature vector;

[0085] (l) Add an 84-bit fully connected layer, a BN layer and a ReLU layer;

[0086] (m) Add a dropout layer with a dropout ratio of 0.2;

[0087] (n) Add a 42-bit fully connected layer, a BN layer and a ReLU layer;

[0088] (o) Add a 16-bit fully connected layer;

[0089] (p) Add an activation layer using the Softmax activation function as the output layer.

[0090] The spatial pyramid pooling is introduced into the network to capture the local details and global information in the image and enhance the feature extraction capability of the network.

[0091] By using a simple network structure and setting a reasonable convolution kernel size, BN layer, Dropout layer, maximum pooling layer and average pooling layer, overfitting is effectively overcome, the training speed is fast, and a higher classification accuracy is achieved.

[0092] Step 5: Lightweight deep learning network training and testing.

[0093] The training set of the embodiment is input into the lightweight deep learning network designed by the present invention. First, forward propagation is performed, and the classification results are output and compared with the true labels of the training set samples. Then, the obtained errors are back-propagated to update and correct the parameters of each layer in the network. The optimizer used for training is SGD, the learning rate is 0.001, the momentum is 0.99, and the batch size is 32.

[0094] After each round of training, the test set of the embodiment is sent to record the classification effect of the network on the test set and compare it with the true labels of the test set samples. It is iterated for 100 rounds, and the training accuracy, training loss, test accuracy and test loss in the entire training and testing process are recorded and the curves are drawn.

[0095] like Figure 5 As shown in the figure, the training loss curve and the test loss curve show a clear downward trend, and the training accuracy curve and the test accuracy curve show a clear upward trend with relatively stable fluctuations. The training loss and the test loss remain stable at a low level after about the 25th epoch, and the training loss is close to 0 at the 40th epoch, indicating that the model converges quickly for the shmoo classification task. In addition, the test accuracy reaches more than 96% at the 60th epoch and remains stable within a smaller fluctuation range, with the highest test accuracy reaching 97.88%. This shows that the model converges quickly for the shmoo classification task, has considerable overfitting control, and has good robustness and classification accuracy.

[0096] Step 6: Complete shmoo classification and calculate various evaluation indicators.

[0097] In this embodiment, the commonly used deep learning networks compared with the lightweight deep learning network include AlexNet, VGG19, ResNet18, MobileNetV2, ShuffleNet, and EfficientNetB0.

[0098] Figure 6 It shows that compared with other networks, the deep learning network designed by the present invention has a faster convergence speed and higher classification accuracy, which not only excellently completes the classification task of the shmoo graph, but also helps to improve the testing efficiency of integrated circuits.

[0099] Supplementary evaluation indicators include macro-average precision, macro-average recall, macro-average F1 score, Matthews correlation coefficient (MCC), network parameter quantity and model size. The table compares the network designed by the present invention with other commonly used networks:

[0100]

[0101] As can be seen from the above table, the parameters and size of the lightweight network designed by the present invention are dozens to hundreds of times smaller than those of other commonly used deep learning networks, but it has achieved obvious advantages in the main evaluation indicators. The macro-average precision, macro-average recall rate and macro-average F1 score reached 98.20%, 97.50% and 97.65% respectively, and the Matthews correlation coefficient reached 0.9762, which is 1% to 5% higher than other networks, proving that the network designed by the present invention is more suitable for shmoo graph classification tasks.

[0102] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A shmoo graph classification method for integrated circuits based on a lightweight deep learning network, characterized in that: The steps include: Step 1: Use ATE to collect the chip's shmoo raw data; Step 2: Preprocess the shmoo raw data to obtain a shmoo image in image format; Step 3: According to the graphic features, the shmoo graph samples obtained in step 2 are classified and labeled to obtain a data set containing 16 shmoo graph categories, and the shmoo data set is divided into a training set and a test set; Step 4: Build a deep learning framework based on lightweight convolutional neural network; Step 5: Send the training set images to the deep learning framework for training. After each round of training, send the test set images, record the classification results of the network for the test set and compare them with the true labels of the test set samples, and repeat this process for multiple rounds. Step 6: Calculate the evaluation index and compare the classification effect of the lightweight deep learning network with other commonly used deep learning networks.

2. A method for classifying integrated circuit shmoo graphs based on a lightweight deep learning network according to claim 1, characterized in that: The specific implementation process of step 1 is as follows: (a) Based on the chip to be tested, a shmoo test plan is developed, which describes how to change the voltage / frequency test conditions at different temperatures; (b) Write the shmoo test program using the ATE software development environment according to the test plan; (c) The chip to be tested is placed in the socket on the load board, and the ATE user host loads and executes the shmoo test program. At the same time, the operating status of the chip under each test condition is recorded and stored in a datalog file in a data format.

3. A method for classifying integrated circuit shmoo graphs based on a lightweight deep learning network according to claim 2, characterized in that: The specific implementation process of step 2 is as follows: (a) Use the data analysis tool to read the datalog file and obtain the shmoo raw data in .csv format with a dimension of 11x11; (b) The data preprocessing program is programmed in Python to convert the shmoo data in .csv format into .png image format according to the Pass / Fail results in each state. The shmoo cells of Pass are green, and the shmoo cells of Fail are red.

4. A method for classifying integrated circuit shmoo graphs based on a lightweight deep learning network according to claim 3, characterized in that: The specific implementation process of step 3 is as follows: (a) The 16 shmoo diagram classification categories are named "Fail-Normal", "Pass-Normal", "Fail-Central", "Pass-Central", "Fail-Hole", "Fail-Central-Hole", "Fail-X-Wall", "Fail-Y-Wall", "Fail-XY-Wall", "Fail-X-Wal" l-Central", "Fail-Y-Wall-Central", "Fail-X-Wall-Hole", "Fail-Y-Wall-Hole", "Fail-XY-Wall-Hole", "Fail-X-Wall-Central-Hole", "Fail-Y-Wall-Central-Hole"; (b) The classification category is composed of four basic features combined with whether it belongs to Pass or Fail. The description of the four basic features is as follows: "Normal" refers to the typical Pass / Fai l situation without other features, and one area in the shmoo diagram is significantly larger than the other; "Central" refers to the Pass / Fai l boundary passing through the central area of ​​the shmoo diagram, and the classification is based on which side has a larger proportion; "Hole" refers to the abnormal scattered points or irregular protrusions / vacancies in the Shmoo diagram, which belongs to Fail; "Wall" refers to the continuous long Fail segments in the X, Y or XY direction under certain test conditions to form a wall-like pattern, which belongs to Fail; (c) The shmoo dataset is divided into a training set and a test set, with the training set and the test set accounting for 70% and 30% of the dataset respectively.

5. A method for classifying integrated circuit shmoo graphs based on a lightweight deep learning network according to claim 4, characterized in that: A deep learning framework based on a lightweight convolutional neural network, which includes an input layer, five convolutional blocks containing convolutional layers, BN layers and ReLU layers, two maximum pooling layers, one average pooling layer, one spatial pyramid pooling layer, two dropout layers, three fully connected layers and a softmax output layer; the specific implementation process is as follows: (a) Add an input layer with an input size of 96x96 and 3 channels; (b) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 128 channels, and a stride of 1; (c) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 64 channels, and a stride of 1; (d) Add a maximum pooling layer with a convolution kernel size of 2x2, 64 channels, and a stride of 2; (e) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 32 channels, and a stride of 1; (f) Add an average pooling layer with a convolution kernel size of 2x2, 32 channels, and a stride of 2; (g) Add a dropout layer with a dropout ratio of 0.2; (h) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 16 channels, and a stride of 1; (i) Add a maximum pooling layer with a convolution kernel size of 2x2, 16 channels, and a stride of 2; (j) Add a convolution block consisting of a convolution layer, a BN layer, and a ReLU layer, with a convolution kernel size of 3x3, 8 channels, and a stride of 1; (k) Add a spatial pyramid pooling layer, use multi-scale pooling with adaptive kernel size, and output a 1x168 feature vector; (l) Add an 84-bit fully connected layer, a BN layer and a ReLU layer; (m) Add a dropout layer with a dropout ratio of 0.2; (n) Add a 42-bit fully connected layer, a BN layer and a ReLU layer; (o) Add a 16-bit fully connected layer; (p) Add an activation layer using the Softmax activation function as the output layer.

6. A method for classifying integrated circuit shmoo graphs based on a lightweight deep learning network according to claim 5, characterized in that: The specific implementation process of step 5 is as follows: (a) After the training set is input into the deep learning network, forward propagation is performed, the classification results are output and compared with the true labels of the training set samples, and then the obtained errors are back-propagated to update and correct the parameters of each layer in the network, thus completing a round of training; (b) After each round of training, the test set images are fed in, and the classification effect of the network on the test set is recorded and compared with the true labels of the test set samples. This is repeated for multiple rounds, and the training accuracy, training loss, test accuracy, and test loss during the entire training and testing process are recorded and plotted.

7. A method for classifying integrated circuit shmoo graphs based on a lightweight deep learning network according to claim 6, characterized in that: The specific implementation process of step 6 is as follows: (a) Plot the test set classification accuracy comparison curve of the lightweight deep learning network and the common deep learning network running on the shmoo dataset; (b) Calculate common evaluation indicators, including macro-average precision, macro-average recall, macro-average F1 score, Matthews correlation coefficient, network parameter number and model size as a comparison supplement.

Images