An unsupervised deep learning method based on synthetic data generation

Abstract

The present invention belongs to the technical field of unsupervised deep learning methods, and specifically relates to an unsupervised deep learning method based on synthetic data generation. A Gaussian mixture model searches for cluster centers and standard deviations of original unlabeled data. Bayesian optimization selects optimal stacking weights for the Gaussian mixture model by measuring the difference between the distribution of data generated by isotropic Gaussian blobs based on the cluster centers and standard deviations of the Gaussian mixture model and the original data. The Gaussian mixture model determines the number of clusters and standard deviations under the optimal stacking weights, and the isotropic Gaussian blobs generate a large amount of labeled and balanced synthetic data based on the number of clusters and standard deviations. The synthetic data is used to train a deep neural network model. The trained deep neural network model performs downstream tasks based on the initial unlabeled data. This method achieves sample balance in the generated synthetic data, assists in the training of the deep neural network, and improves model accuracy in unsupervised learning scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of unsupervised deep learning methods, and in particular relates to an unsupervised deep learning method based on synthetic data generation. Background Art

[0002] With the development of wireless networks, mining data information and extracting its features for application have become common tasks in the data age. Examples include unlabeled image classification in computer vision and abnormal traffic detection in networks. However, in the real world, a wide range of unlabeled data exists. This data, lacking labels to serve as a reference and guide for training, poses significant challenges for data mining. Currently, the core idea of common methods for data mining on unlabeled data is to analyze the statistical characteristics and similarities of the data to discover underlying structures and patterns. This approach is known as unsupervised training. Deep neural networks stand out among these methods due to their flexible adaptability. They can implement complex feature transformations, adapting to diverse unsupervised learning tasks.

[0003] Deep neural networks require a large amount of data as input to extract and summarize patterns in data. The current mainstream solution is to generate synthetic data based on existing samples, thereby enriching the sample size involved in training. Synthetic data generation follows the paradigm of first analyzing the structure of the original data and then generating synthetic data based on the analysis results. Examples include data generation methods based on principal component analysis and those based on Gaussian mixture functions. While these methods can address the limited sample size issue in unsupervised deep learning to a certain extent, they lack attention to sample balance in synthetic data generation. For example, the number of samples of different types is uneven, and the feature differentiation between different sample types is not high. This leads to confusion in the training process of deep neural networks, resulting in poor results in practical applications and making it difficult to meet the needs of practical applications.

[0004] Therefore, how to improve the training process of deep neural networks in existing methods to achieve sample balance in generating synthetic data, assist in the training of deep neural networks, and improve the model accuracy in unsupervised learning scenarios is a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The purpose of the present invention is to provide an unsupervised deep learning method based on synthetic data generation, so as to achieve sample balance in the generated synthetic data, assist in the training of deep neural networks, and improve the model accuracy in unsupervised learning scenarios.

[0006] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0007] An unsupervised deep learning method based on synthetic data generation, comprising the following steps:

[0008] S1: The Gaussian mixture model obtains the superposition weights from the Bayesian optimizer and finds the cluster centers and standard deviations based on the unlabeled original data. The isotropic Gaussian spots generate synthetic data using the cluster centers and standard deviations of the current Gaussian mixture model.

[0009] S2: measuring the distribution difference between the synthetic data and the original data, and obtaining the optimal superposition weight for the Gaussian mixture function based on Bayesian optimization iteration;

[0010] S3: Isotropic Gaussian uses the cluster center and standard deviation of the Gaussian mixture model under the optimal superposition weight as the benchmark as a synthetic data generator. The synthetic data generator generates labeled synthetic data with a distribution similar to the original data.

[0011] S4: Inputting the labeled synthetic data into a deep neural network model for model training;

[0012] S5: Input the initial unlabeled data into the trained deep neural network model to perform downstream tasks.

[0013] Preferably, the specific process of finding the cluster center and standard deviation based on the unlabeled original data in step S1 is as follows:

[0014] S11: The unlabeled raw data is represented as X ,Include N samples, each sample i Include M Features, any x i,j Indicates the i The first sample j The value of the feature in , where 1≤ i ≤ N , 1≤ j ≤ M , the unlabeled original data X Divide K kind, K =[1, 2, ···, k ];

[0015] S11: The unlabeled raw data is represented as X ,Include N samples, each sample i Include M Features, any x i,j Indicates the i The first sample j The value of the feature in , where 1≤ i ≤ N , 1≤ j≤ M , the unlabeled original data X Divide K kind, K =[1, 2, ···, k ];

[0016] S12: Unlabeled raw data X Perform clustering processing, the specific formula is as follows:

[0017] ;

[0018] in, is the clustering result, For samples i The class to which it belongs, , is the class to which sample 1 belongs, is the class to which sample 2 belongs, For samples N The class to which it belongs;

[0019] S13: Calculate the cluster center based on the probability of the given cluster calculation data sample C and standard deviation σ .

[0020] Preferably, in step S13, the cluster center is calculated based on the probability of the given cluster calculation data sample. C The specific process of calculating the value of the sum of the standard deviation σ is as follows:

[0021] S131: Based on K The Gaussian mixture model is created by linear superposition of Gaussian models. The specific formula is as follows:

[0022] ;

[0023] in, w k is the superposition parameter, , For the k The mean vector of the Gaussian components, For the k The covariance matrix of the Gaussian components, For the k The multivariate normal density function of the components;

[0024] S132: Bayesian optimizer randomly initializes superposition parameters for Gaussian mixture models w , and calculate the original sample mean based on the original data , the covariance matrix , generating the sample mean , then the Gaussian mixture model has 3×k Parameters to be updated;

[0025] S133: Calculate each sample i The probability of belonging to each cluster , a total of iterative calculations n × k The calculation formula is as follows:

[0026] ;

[0027] in, It is j The mixing weights of the components, , x represents the sample set, x i For the i samples, For the j The mean vector of the Gaussian components, For the j The covariance matrix of the Gaussian components;

[0028] S134: Isotropic Gaussian blobs generate labeled and balanced synthetic data of the same scale as the original data.

[0029] Preferably, in each iteration in step S133, the parameters are updated using the following formula:

[0030] ;

[0031] ;

[0032] ;

[0033] in, is the sample mean, , N is the total number of data samples, I is the identity matrix, is the regularization term.

[0034] Preferably, the specific process of generating labeled and balanced synthetic data of the same scale as the original data by isotropic Gaussian spots in step S134 is as follows:

[0035] S1341: Isotropic Gaussian blobs generate labels based on the optimal weighted parameters of the Gaussian mixture model: K Randomly select a cluster center from the clusters k , at this time the data object d i belong k , the generator is based on the labeled data objectd i To generate labels, d i ∈ D , and get a size of N The included dataset D Tag collection ;

[0036] S1342: Data generator isotropic Gaussian spots according to C k , σ k is the center and standard deviation, For the sample size, generate a data object d i , the formula is as follows:

[0037] ;

[0038] in, Normal distribution , the generation process is executed K times until the isotropic Gaussian spots generate samples for all clusters, and the downstream obtains a N Balanced dataset of synthetic samples D .

[0039] Preferably, the specific process of obtaining the optimal superposition weight for the Gaussian mixture function based on Bayesian optimization iteration in step S2 is as follows:

[0040] S21: The Kolmogorov-Smirnov test is used to determine whether the data generated by the generator is similar to the original data distribution. The statistic S of the Kolmogorov-Smirnov test is:

[0041] S= sup x |F m ( x )- G n ( x ) | ;

[0042] in, sup x express |F m ( x )- G n ( x ) | Maximum value, F m ( x ) is the empirical distribution function of the original data,G n ( x ) is the empirical distribution function of the synthetic data,

[0043] ;

[0044] ;

[0045] I is the indicator function, when X i ≤ X , Y i ≤ X When I=1, n is the sample size, m represents the number of original data samples;

[0046] S22: Loop through steps S12, S13, and S21 to obtain the optimal superposition weight of the Gaussian mixture model.

[0047] Preferably, the specific process of generating labeled synthetic data in step S3 is as follows:

[0048] S31: Based on the obtained optimal superposition weight of the Gaussian mixture model, run the Gaussian mixture function to obtain the cluster center obtained by iterative Gaussian mixture model under this weight C and standard deviation σ;

[0049] S32: Specify the required number of samples, run the isotropic Gaussian spot algorithm in rounds according to the number of cluster centers, and generate labeled sample-balanced synthetic data D'.

[0050] Preferably, the specific process of inputting the labeled synthetic data into the deep neural network model for model training in step S4 is as follows:

[0051] S41: Deploy a deep neural network model, including a mask layer, a dense layer, a batch normalization layer, and a random dropout layer. The mask layer is the first layer of the deep neural network model, and the dense layer has multiple neurons and uses a linear transformation and an activation function that increases nonlinearity.

[0052] S42: Using the mask layer, the median of the input sequence is compared with the preset specified value. Equal samples are masked, the specific formula is as follows:

[0053] ;

[0054] in, represents the i-th element of the mask vector, Represents the i-th sample value in the input sequence;

[0055] S43: Calculation is performed through dense layers. The specific formula is as follows:

[0056] h=ReLU(WD'+b);

[0057] Among them, h is the output after dense layer calculation, W is the linearly changing learnable weight matrix, b is the bias vector, ReLU is the rectified linear unit function, ReLU= max (0, z ), z It is the result of linear transformation in the neural network. When the linear unit function is corrected by ReLU, if z>0, the output is z ; Otherwise, output 0, and D' is the synthetic data balanced with labeled samples;

[0058] S44: Input the output result of the dense layer to the batch normalization layer for processing. The specific formula is as follows:

[0059] ;

[0060] in, and They are h The mean and standard deviation of the current batch, and are learnable parameters;

[0061] S45: The output of the batch normalization layer is input to the random inactivation layer for processing. The specific formula is as follows:

[0062] ;

[0063] in, d is a binary masking vector, and the probability of each element being 0 is p , the probability of each element being 1 is 1 / p ;

[0064] S46: Output of the random dropout layer Enter with k The output dense layer of neurons and softmax activation function calculates the probability that the input data belongs to each class of labels:

[0065] ;

[0066] The dimension of y is k ×N, is the weight matrix, is the bias vector, and the softmax activation function is:

[0067] ;

[0068] in, zi is the input vector z The i elements;

[0069] S47: The cross entropy is used to measure the probability of each type of predicted label:

[0070] ;

[0071] in, Indicates the class to which sample i belongs, synlog indicates the synchronous logarithm, To predict the probability, the training loss is used to update the learnable parameters of the deep neural network through the Adam optimizer until the deep neural network converges.

[0072] Preferably, the specific process of step S5 is as follows:

[0073] The initial unlabeled data samples X Input the deep neural network model pre-trained by the synthetic data in step S4, output the corresponding results, and execute downstream tasks.

[0074] The beneficial effects of the present invention include:

[0075] The present invention provides an unsupervised deep learning method based on synthetic data generation. The Gaussian mixture model searches for the cluster centers and standard deviations of the original unlabeled data. Bayesian optimization selects the optimal superposition weight for the Gaussian mixture model by measuring the difference between the distribution of data generated by isotropic Gaussian blobs based on the cluster centers and standard deviations of the Gaussian mixture model and the original data. The Gaussian mixture model determines the number of clusters and the standard deviation under the optimal superposition weight. The isotropic Gaussian blobs generate a large amount of labeled and balanced synthetic data based on the number of clusters and the standard deviation. The synthetic data is used to train a deep neural network model. The trained deep neural network model performs downstream tasks based on the initial unlabeled data. The method can generate a large amount of labeled synthetic data with a similar distribution to the unlabeled data based on the statistical characteristics of the unlabeled data, thereby providing sufficient training samples for the deep neural network model training in an unsupervised scenario, improving the performance of the model, achieving sample balance in generating synthetic data, assisting the training of the deep neural network, and improving the model accuracy in an unsupervised learning scenario.

[0076] The present invention mines the statistical characteristics of unlabeled data by superimposing weights on a Bayesian optimized Gaussian mixture model, generates labeled synthetic data to expand the training sample size, and provides a label reference for deep neural network model training, thereby improving model performance. Considering the sample size of each type of synthetic data, a consistent number of samples is generated for each type of data to achieve balanced synthetic data samples and avoid confusion caused by unbalanced generated samples in deep neural network model learning. The structure of the deep neural network can be flexibly adjusted according to specific tasks, and the output of the pre-trained deep neural network can be further processed to meet the needs of various downstream tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 Schematic diagram of the process of an unsupervised deep learning method based on synthetic data generation according to the present invention.

[0078] Figure 2 This is a schematic diagram of the process of selecting the optimal superposition weights for the Gaussian mixture model by the Bayesian optimizer of the present invention. DETAILED DESCRIPTION

[0079] The following is combined with Figure 1~Figure 2 The present invention is described in further detail:

[0080] Example 1

[0081] See also Figure 1 and Figure 2 As shown, an unsupervised deep learning method based on synthetic data generation includes the following steps:

[0082] S1: The Gaussian mixture model randomly obtains superposition weights from the Bayesian optimizer and finds cluster centers and standard deviations based on the unlabeled original data. The isotropic Gaussian spot generates synthetic data using the cluster centers and standard deviations of the current Gaussian mixture model.

[0083] S2: Kolmogorov-Smirnov test measures the distribution difference between the synthetic data and the original data, and obtains the optimal superposition weight for the Gaussian mixture function based on Bayesian optimization iteration;

[0084] S3: Isotropic Gaussian uses the cluster centers and standard deviations of the Gaussian mixture model under the optimal superposition weights as the benchmark as a synthetic data generator. The synthetic data generator generates labeled synthetic data with a distribution similar to the original data.

[0085] Run the Gaussian mixture function with the optimal parameters of the Gaussian mixture model, and then specify the required number of samples N' to run the isotropic Gaussian spot generation label and synthetic data. To ensure the balance of the generated data, the generator will call the cluster center and variance one by one to generate synthetic data with a balanced sample size of each type. The isotropic Gaussian spots are called K times. Each time a C k generate samples, of which , the downstream gets a balanced dataset of size , the clusters are represented as labels for each synthetic data.

[0086] S4: Input the labeled synthetic data into a deep neural network model for model training. In this step, an adjustable deep neural network model is designed to adapt to the execution needs of downstream tasks. The deep neural network includes mask layers, dense layers, batch normalization layers, and random dropout layers.

[0087] The combination of these network layers can be flexibly adjusted for different specific tasks. The output is ultimately fed into an output dense layer with k neurons and a softmax activation function, which calculates the probability of the input data belonging to each label. The predicted label probabilities are measured using cross-entropy as the training loss, and the learnable parameters of the deep neural network model are updated using gradient descent until the deep neural network converges.

[0088] S5: Input the initial unlabeled data into the trained deep neural network model, perform downstream tasks, and output the execution results.

[0089] Example 2

[0090] Based on Example 1, the specific process of finding the cluster center and standard deviation based on the unlabeled original data in step S1 is as follows:

[0091] S11: Constructing structured data representation: Representing unlabeled raw data as X ,Include N samples, each sample i Include M Features, any x i,j Indicates the i The first sample j The value of the feature in , where 1≤ i ≤ N , 1≤ j ≤ M , the unlabeled original data X Divide K kind, K =[1, 2, ···, k ];

[0092] S12: Unlabeled raw data X Perform clustering processing, the specific formula is as follows:

[0093] ;

[0094] in, is the clustering result, For samples i The class to which it belongs, , is the class to which sample 1 belongs, is the class to which sample 2 belongs, For samples N The class to which it belongs;

[0095] S13: Calculate the cluster center based on the probability of the given cluster calculation data sample C and standard deviation σ ;

[0096] S13: Calculate the cluster center based on the probability of the given cluster calculation data sample C and standard deviation σ.

[0097] In this embodiment, in step S13, the cluster center is calculated based on the probability of the given cluster calculation data sample. C The specific process of calculating the value of the sum of the standard deviation σ is as follows:

[0098] S131: Based on K The Gaussian mixture model is created by linear superposition of Gaussian models. The specific formula is as follows:

[0099] ;

[0100] in, w k is the superposition parameter, , For the k The mean vector of the Gaussian components, For the k The covariance matrix of the Gaussian components, For the k The multivariate normal density function of the components;

[0101] S132: Bayesian optimizer randomly initializes superposition parameters for Gaussian mixture models w , and calculate the original sample mean based on the original data , the covariance matrix , generating the sample mean , then the Gaussian mixture model has 3× k Parameters to be updated;

[0102] S133: Calculate each sample i The probability of belonging to each cluster , a total of iterative calculations n ×k The calculation formula is as follows:

[0103] ;

[0104] in, It is j The mixing weights of the components, , x represents the sample set, x i For the i samples, For the j The mean vector of the Gaussian components, For the j The covariance matrix of the Gaussian components;

[0105] S134: Isotropic Gaussian blobs generate labeled and balanced synthetic data of the same scale as the original data.

[0106] In each iteration in step S133, the parameters are updated using the following formula:

[0107] ;

[0108] ;

[0109] ;

[0110] in, is the sample mean, , N is the total number of data samples, I is the identity matrix, is the regularization term.

[0111] The specific process of generating labeled and balanced synthetic data of the same scale as the original data by isotropic Gaussian spots in step S134 is as follows:

[0112] S1341: Isotropic Gaussian blobs generate labels based on the optimal weighted parameters of the Gaussian mixture model: K Randomly select a cluster center from the clusters k , at this time the data object d i belong k , the generator is based on the labeled data object d i To generate labels, d i ∈ D , and get a size of N The included dataset D Tag collection ;

[0113] S1342: Data generator isotropic Gaussian spots according to C k , σ k is the center and standard deviation, For the sample size, generate a data object d i , the formula is as follows:

[0114] ;

[0115] in, Normal distribution , the generation process is executed K times until the isotropic Gaussian spots generate samples for all clusters, and the downstream obtains a N Balanced dataset of synthetic samples D .

[0116] Isotropic Gaussian spots generate synthetic data of the same size as the original data. At the same time, the present invention focuses on the balance of the generated synthetic data. Therefore, in the data generation process, isotropic Gaussian spots are called in turn based on the number of samples and the number of clusters to generate data with balanced sample size. Isotropic Gaussian spots use the cluster center C and standard deviation σ generated in the Gaussian mixture function as the cluster center and standard deviation, where C k , σ k Representing the cluster center and standard deviation of a specific class, respectively, a dataset with similar characteristics and the same size as the original data is generated. To ensure the sample balance of the synthetic data, the isotropic Gaussian spot is called cyclically.

[0117] Example 3

[0118] On the basis of Example 1 or Example 2, the specific process of obtaining the optimal superposition weight for the Gaussian mixture function based on Bayesian optimization iteration in step S2 is as follows:

[0119] S21: The Kolmogorov-Smirnov test is used to determine whether the data generated by the generator is similar to the original data distribution. The statistic S of the Kolmogorov-Smirnov test is:

[0120] S= sup x |F m ( x )- G n ( x ) | ;

[0121] in, sup x express |F m ( x )- G n ( x ) | Maximum value.

[0122] F m ( x ) is the empirical distribution function of the original data, G n ( x ) is the empirical distribution function of the synthetic data,

[0123] ;

[0124] ;

[0125] I is the indicator function, when X i ≤ X , Y i ≤ X When , I=1, otherwise it is 0, n is the sample size, m Indicates the number of original data samples;

[0126] S22: Loop through steps S12, S13, and S21 to obtain the optimal superposition weight of the Gaussian mixture model.

[0127] The specific process of generating labeled synthetic data in step S3 is as follows:

[0128] S31: Based on the optimal superposition weight of the Gaussian mixture model obtained, run the Gaussian mixture function to obtain the cluster center C and standard deviation obtained by the iteration of the Gaussian mixture model under this weight σ ;

[0129] S32: Specify the required number of samples and run isotropic Gaussian spots according to the number of cluster centers to generate labeled sample balanced synthetic data .

[0130] Example 4

[0131] On the basis of Example 1 or Example 2 or Example 3, the specific process of inputting the labeled synthetic data into the deep neural network model for model training in step S4 is as follows:

[0132] S41: Deploy a deep neural network model, including a mask layer, a dense layer, a batch normalization layer, and a random dropout layer. The mask layer is the first layer of the deep neural network model, and the dense layer has multiple neurons and uses linear transformation and a ReLU activation function to increase nonlinearity. The combination of the above network layers can be flexibly adjusted for different specific tasks.

[0133] S42: Compare the median of the input sequence with the preset specified value through the mask layer Equal samples are masked, the specific formula is as follows:

[0134] ;

[0135] represents the i-th element of the mask vector, Represents the i-th sample value in the input sequence.

[0136] S43: Calculation is performed through dense layers. The specific formula is as follows:

[0137] h=ReLU(WD'+b);

[0138] Among them, h is the output after dense layer calculation, W is the linearly changing learnable weight matrix, b is the bias vector, ReLU is the rectified linear unit function, ReLU= max (0, z ), z It is the result of linear transformation in the neural network. When the linear unit function is corrected by ReLU, if z>0, the output is z ; Otherwise, output 0, and D' is the synthetic data balanced with labeled samples.

[0139] S44: Input the output result of the dense layer to the batch normalization layer for processing. The normalization layer can disrupt the data order, accelerate the convergence of the network, and improve the generalization ability of the network. The specific formula is as follows:

[0140] ;

[0141] in, and They are h The mean and standard deviation of the current batch, and are learnable parameters;

[0142] S45: The output of the batch normalization layer is input to the random activation layer for processing. The random activation layer can avoid the problem of overfitting of deep neural networks and reduce training time. The specific formula is as follows:

[0143] ;

[0144] in, d is a binary masking vector, and the probability of each element being 0 is p , the probability of each element being 1 is 1 / p ;

[0145] S46: Predicting labels based on synthetic data: Output of the dropout layer Enter with k The output dense layer of neurons and softmax activation function calculates the probability that the input data belongs to each class of labels:

[0146] ;

[0147] The dimension of y is k ×N, is the weight matrix, is the bias vector, and the softmax activation function is:

[0148] ;

[0149] Where, is the input vector z The i elements, z j is the input vector z The j elements;

[0150] S47: Backpropagation training of deep neural networks: The probability results of the predicted labels are measured using cross entropy:

[0151] ;

[0152] in, Indicates the class to which sample i belongs, synlog indicates the synchronous logarithm, To predict the probability, the training loss is used to update the learnable parameters of the deep neural network through the Adam optimizer until the deep neural network converges.

[0153] The specific process of step S5 is as follows:

[0154] The initial unlabeled data samples X Input the deep neural network model pre-trained by the synthetic data in step S4, output the corresponding results, and execute downstream tasks.

[0155] The present invention mines the statistical characteristics of unlabeled data by superimposing weights on a Bayesian optimized Gaussian mixture model, generates labeled synthetic data to expand the training sample size, and provides a label reference for deep neural network model training, thereby improving model performance. Considering the sample size of each type of synthetic data, a consistent number of samples is generated for each type of data to achieve balanced synthetic data samples and avoid confusion caused by imbalanced generated samples in deep neural network model learning. The structure of the deep neural network of the present invention can be flexibly adjusted according to specific tasks, and the output of the pre-trained deep neural network supports further processing, which can meet the needs of various downstream tasks. It can be applied to a variety of downstream tasks and supports setting other data processing methods for specific problems based on the results of pre-training.

[0156] In summary, the present invention provides an unsupervised deep learning method based on synthetic data generation. The Gaussian mixture model finds the cluster centers and standard deviations of the original unlabeled data. Bayesian optimization selects the optimal superposition weight for the Gaussian mixture model by measuring the difference between the distribution of the isotropic Gaussian blobs generated based on the cluster centers and standard deviations of the Gaussian mixture model and the original data. The Gaussian mixture model determines the number of clusters and standard deviation under the optimal superposition weights. The isotropic Gaussian blobs generate a large amount of labeled and balanced synthetic data based on the number of clusters and standard deviation. The synthetic data is used to train a deep neural network model. The trained deep neural network model performs downstream tasks based on the initial unlabeled data. Based on the statistical characteristics of the unlabeled data, a large amount of labeled synthetic data with a similar distribution can be generated based on the clustering method, thereby providing sufficient training samples for the deep neural network model training in the unsupervised scenario and improving the performance of the model. By analyzing the distribution characteristics of the original data, the clustering method is used to generate synthetic data that matches the original data distribution and is balanced to assist the training of the deep neural network, thereby improving the model accuracy in the unsupervised learning scenario. This effectively solves the technical problem that the current unbalanced generation of synthetic data leads to low accuracy of the deep learning model.

Claims

1. An unsupervised deep learning method based on synthetic image data generation, characterized in that The following steps are involved: S1: In the unlabeled image classification task in computer vision, the Gaussian mixture model obtains superposition weights from the Bayesian optimizer, finds cluster centers and standard deviations based on the unlabeled original image data, and the isotropic Gaussian spots use the cluster centers and standard deviations of the current Gaussian mixture model to generate synthetic image data; S2: measuring the distribution difference between the synthesized image data and the original image data, and obtaining an optimal superposition weight for the Gaussian mixture function based on Bayesian optimization iteration; S3: Isotropic Gaussian takes the cluster center and standard deviation of the Gaussian mixture model under the optimal superposition weight as the benchmark, and is used as a synthetic image data generator to generate labeled synthetic image data with a distribution similar to the original image data; S4: Inputting the labeled synthetic image data into a deep neural network model for model training; S5: Input the initial unlabeled image data into the trained deep neural network model to perform downstream tasks; The specific process of finding the cluster center and standard deviation based on the unlabeled original image data in step S1 is as follows: S11: The unlabeled raw image data is represented as X ,Include N samples, each sample i Include M Features, any x i,j Indicates the i The first sample j The value of the feature in , where 1≤ i ≤ N , 1≤ j ≤ M , the unlabeled original image data X Divide K kind, K =[1, 2, ···, k ]; S12: Unlabeled original image data X Perform clustering processing, the specific formula is as follows: ； in, is the clustering result, For samples i The class to which it belongs, , is the class to which sample 1 belongs, is the class to which sample 2 belongs, For samples N The class to which it belongs; S13: Calculate the cluster center based on the probability of the given clustering image data sample C and standard deviation σ; In step S13, the cluster center is calculated based on the probability of the given clustering image data sample. C The specific process of calculating the value of the sum of the standard deviation σ is as follows: S131: Based on K The Gaussian mixture model is created by linear superposition of Gaussian models. The specific formula is as follows: ； in, w k is the superposition parameter, , For the k The mean vector of the Gaussian components, For the k The covariance matrix of the Gaussian components, For the k The multivariate normal density function of the components; S132: Bayesian optimizer randomly initializes superposition parameters for Gaussian mixture models w , and calculate the original sample mean based on the original image data , the covariance matrix , generating the sample mean , then the Gaussian mixture model has 3× k Parameters to be updated; S133: Calculate each sample i The probability of belonging to each cluster , a total of iterative calculations n × k The calculation formula is as follows: ； in, It is j The mixing weights of the components, , x represents the sample set, x i For the i samples, For the j The mean vector of the Gaussian components, For the j The covariance matrix of the Gaussian components; S134: Generate labeled and balanced synthetic image data of the same scale as the original image data using isotropic Gaussian spots; The specific process of generating labeled and balanced synthetic image data of the same scale as the original image data by isotropic Gaussian spots in step S134 is as follows: S1341: Isotropic Gaussian blobs generate labels based on the optimal weighted parameters of the Gaussian mixture model: K Randomly select a cluster center from the clusters k , at this time the image data object d i belong k , the generator is based on the labeled image data object d i To generate labels, d i ∈ D , and get a size of N The image dataset D Tag collection ; S1342: Image data generator isotropic Gaussian spots according to C k , σ k is the center and standard deviation, For the sample size, generate image data objects d i , the formula is as follows: ； in, Normal distribution , the generation process is executed K times until the isotropic Gaussian spots generate samples for all clusters, and the downstream obtains a N A balanced image dataset of synthetic samples D .

2. The unsupervised deep learning method based on synthetic image data generation according to claim 1, characterized in that: In each iteration in step S133, the parameters are updated using the following formula: ； ； ； in, is the sample mean, , N is the total number of image data samples, I is the identity matrix, is the regularization term.

3. The unsupervised deep learning method based on synthetic image data generation according to claim 1, characterized in that: The specific process of obtaining the optimal superposition weight for the Gaussian mixture function based on Bayesian optimization iteration in step S2 is as follows: S21: The Kolmogorov-Smirnov test is used to determine whether the image data generated by the generator is similar to the original image data distribution. The statistic S of the Kolmogorov-Smirnov test is: S= sup x |F m ( x )- G n ( x ) | ； in, sup x express |F m ( x )- G n ( x ) | Maximum value, F m ( x ) is the empirical distribution function of the original image data, G n ( x ) is the empirical distribution function of the synthetic image data, ； ； I is the indicator function, when X i ≤ X , Y i ≤ X When I=1, n is the sample size, m represents the number of original image data samples; S22: Loop through steps S12, S13, and S21 to obtain the optimal superposition weight of the Gaussian mixture model.

4. The unsupervised deep learning method based on synthetic image data generation according to claim 3, characterized in that: The specific process of generating labeled synthetic image data in step S3 is as follows: S31: Based on the obtained optimal superposition weight of the Gaussian mixture model, run the Gaussian mixture function to obtain the cluster center obtained by iterative Gaussian mixture model under this weight C and standard deviation σ ; S32: Specify the required number of samples, run the isotropic Gaussian spot method in rounds according to the number of cluster centers, and generate labeled sample-balanced synthetic image data D'.

5. The unsupervised deep learning method based on synthetic image data generation according to claim 1, characterized in that: The specific process of inputting the labeled synthetic image data into the deep neural network model for model training in step S4 is as follows: S41: Deploy a deep neural network model, including a mask layer, a dense layer, a batch normalization layer, and a random dropout layer. The mask layer is the first layer of the deep neural network model, and the dense layer has multiple neurons and uses a linear transformation and an activation function that increases nonlinearity. S42: Using the mask layer, the median of the input sequence is compared with the preset specified value. Equal samples are masked, the specific formula is as follows: ； in, represents the i-th element of the mask vector, Represents the i-th sample value in the input sequence; S43: Calculation is performed through dense layers. The specific formula is as follows: h=ReLU(WD'+b); Among them, h is the output after dense layer calculation, W is the linearly changing learnable weight matrix, b is the bias vector, ReLU is the rectified linear unit function, ReLU= max (0, z ), z It is the result of linear transformation in the neural network. When the linear unit function is modified by ReLU, if z>0, the output is z ; Otherwise, output 0, D' is the synthetic image data balanced by labeled samples; S44: Input the output result of the dense layer to the batch normalization layer for processing. The specific formula is as follows: ； in, and They are h The mean and standard deviation of the current batch, and are learnable parameters; S45: The output of the batch normalization layer is input to the random inactivation layer for processing. The specific formula is as follows: ； in, d is a binary masking vector, and the probability of each element being 0 is p , the probability of each element being 1 is 1 / p ; S46: Output of the random dropout layer Enter with k The output dense layer of neurons and softmax activation function calculates the probability that the input image data belongs to each class of labels: ； The dimension of y is k ×N, is the weight matrix, is the bias vector, and the softmax activation function is: ； in, z i is the input vector z The i elements; S47: The cross entropy is used to measure the probability of each type of predicted label: ； in, Indicates the class to which sample i belongs, synlog indicates the synchronous logarithm, To predict the probability, the training loss is used to update the learnable parameters of the deep neural network through the Adam optimizer until the deep neural network converges.

6. The unsupervised deep learning method based on synthetic image data generation according to claim 1, characterized in that: The specific process of step S5 is as follows: The initial unlabeled image data samples X Input the deep neural network model pre-trained by the synthetic image data in step S4, output the corresponding results, and execute downstream tasks.

Images