Generative adversarial network data expansion method based on topological self-organizing structure

Abstract

The invention discloses a generative adversarial network data expansion method based on a topological self-organizing structure, and the method comprises the steps: carrying out the operation of a generator and a discriminator of a data expansion model in a conventional adversarial training mode at an initial stage; after a preset round of iteration is carried out, a nonlinear feature entanglement decoupling strategy is adopted, the features are recoded, and the linear relation between the original features is broken, so that a more complex and changeable data sample is generated; after the generated data sample is evaluated by the discriminator, an evaluation result is fed back to the generator; the generator adjusts the parameters according to the feedback, and optimizes the quality and diversity of the generated samples; a progressive training strategy is adopted to further improve the quality of the generated data sample, and low-dimensional features are sequentially transited to high-dimensional features; and the steps are repeated and iterated. According to the method, the problem of insufficient training samples is effectively solved, the samples with more complex distribution can be generated, and the diversity and quality of the samples are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a generative adversarial network data expansion method based on a topological self-organizing structure. Background Art

[0002] At present, there are still many problems in intelligent dialogue systems and related technical fields, which make it difficult to meet the increasingly complex user needs and application scenarios. With the rapid development of artificial intelligence technology, intelligent dialogue systems are widely used in customer service, education, medical care, finance and other fields. For example, the digitalization and intelligentization process in the field of industrial equipment continues to advance, and the automated operation, remote monitoring and fault diagnosis of equipment have become industry trends. However, the data of industrial equipment such as operation manuals, maintenance records, fault logs, etc. are numerous and complex, involving not only real-time data of machine operation, but also a large amount of natural language text processing, such as equipment fault reports, technical documents, etc. How to make the model effectively understand this data and provide intelligent decision support and fault solutions has become a major challenge in current technology.

[0003] However, in many specific fields, especially in the field of industrial equipment, the cost of acquiring training data is high, and insufficient data often leads to poor model generalization ability. Traditional data expansion methods, such as data enhancement and simple GAN models, are difficult to generate samples with sufficient complexity and diversity. Insufficient training samples can easily lead to poor model generalization ability and affect the accuracy of the model, making it difficult to adapt to the natural language processing needs related to industrial equipment.

[0004] In view of this, it is necessary to provide a generative adversarial network data expansion method based on topological self-organizing structure to overcome the above defects. Summary of the invention

[0005] The purpose of the present invention is to provide a data expansion method for a generative adversarial network based on a topological self-organizing structure, aiming to solve the problem that traditional data expansion methods are difficult to generate samples with sufficient complexity and diversity.

[0006] In order to achieve the above object, the present invention provides a data expansion method for a generative adversarial network based on a topological self-organizing structure, which is applied to a data expansion model based on a generative adversarial network, comprising:

[0007] Step S11: In the initial stage, the generator and discriminator of the data augmentation model are trained in a traditional adversarial manner. The generator tries to generate samples that are sufficient to deceive the discriminator, and the discriminator tries to distinguish between real samples and generated samples. Multiple rounds of iterations are performed to gradually improve the capabilities of the generator and the discriminator.

[0008] Step S12: After a preset number of iterations, a nonlinear feature entanglement decoupling strategy is used to re-encode the features and break the linear relationship between the original features to generate more complex and varied data samples;

[0009] Step S13: After the generated data samples are evaluated by the discriminator, the evaluation results are fed back to the generator; the generator adjusts its parameters according to the feedback to optimize the quality and diversity of the generated samples;

[0010] Step S14: using a progressive training strategy to further improve the quality of generated data samples, transitioning from low-dimensional features to high-dimensional features;

[0011] Step S15: Repeat the above steps S12-S14 until the preset stop iteration condition is met, which means that the data expansion model training is completed.

[0012] In a preferred embodiment, in step S11, the initial adversarial training step is expressed as:

[0013] G(z;θ g )=x fake ,

[0014] D(x;θ d ) = Sig(x),

[0015] In the formula, G represents the generator, and z is the output from the prior distribution p z The random noise vector extracted from g is the parameter of the generator; x fake is the fake sample output by the generator; D represents the discriminator, x is the input sample, θ d are the parameters of the discriminator, and Sig() is the Sigmoid activation function.

[0016] In a preferred embodiment, in step S12, the nonlinear characteristic entanglement decoupling strategy is expressed as:

[0017] x ent =W ent ·x fake +b ent ,

[0018] x dis =ReLU(x ent ),

[0019] Where W ent and b ent are the weights and biases of the feature entanglement layer respectively; ReLU() is the linear rectification function; x ent is the intermediate feature vector after feature entanglement; x disis the decoupled eigenvector.

[0020] In a preferred embodiment, the weight W of the feature entanglement layer ent and bias b ent The calculation method is expressed as:

[0021] W ent =UΣV T ,

[0022] b ent =μ x -W ent μ z ,

[0023] Where U, Σ and V T Obtained by performing singular value decomposition on the covariance matrix of the generated samples; μ x is the mean of the input features, μ z is the mean of the noise vector z.

[0024] In a preferred embodiment, in step S13, the feedback adjustment method is expressed as:

[0025] loss d = -logD(x real )-log(1-D(G(z))),

[0026] loss g = -logD(G(z)),

[0027]

[0028] In the formula, loss d is the loss function of the discriminator, loss g is the loss function of the generator; x real is a real sample; α d and α g are the learning rates of the discriminator and the generator respectively; and represents the gradient of the loss function with respect to their respective parameters.

[0029] In a preferred embodiment, the learning rate α d and α g It is set by dynamic adjustment, and the adjustment method is expressed as:

[0030]

[0031] In the formula, α d0 and α g0 are the initial learning rates of the discriminator and the generator respectively; β dand β g is the learning rate decay coefficient, and t is the number of training rounds.

[0032] In a preferred embodiment, α d0 and α g0 Set to 0.03 and 0.01 respectively.

[0033] In a preferred embodiment, in step S14, the progressive training strategy is expressed as:

[0034] x prog =λ prog ·x dis +(1-λ prog )·x real ,

[0035] In the formula, x prog is the sample generated by progressive training; prog It is a coefficient that controls the fusion ratio of real samples and generated samples.

[0036] In a preferred embodiment, the coefficient λ of the fusion ratio prog The calculation method is expressed as:

[0037]

[0038] In the formula, γ c is a parameter that controls the convergence speed of the coefficients.

[0039] In a preferred embodiment, γ c Set to 3.

[0040] The present invention provides a topological self-organizing structure-based generative adversarial network data expansion method, which uses a topological self-organizing structure-based generative adversarial network for data expansion, effectively solving the problem of insufficient training samples, and breaking the linear correlation in the generated samples through a nonlinear feature entanglement decoupling strategy, thereby generating samples with more complex distribution, improving the diversity and quality of the samples, and enhancing the generalization ability of the model for subsequent training of the expanded samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 A flow chart of a method for expanding data in a generative adversarial network based on a topological self-organizing structure provided by the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and beneficial technical effect of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. It should be understood that the specific implementation methods described in this specification are only for explaining the present invention, not for limiting the present invention.

[0044] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0045] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0046] In an embodiment of the present invention, a method for data expansion of a generative adversarial network based on a topological self-organizing structure is provided, which is applied to a data expansion model based on a generative adversarial network. Aiming at the deficiencies of traditional generative adversarial networks in terms of sample diversity and training stability, the present invention adopts a nonlinear feature entanglement strategy in the generator, and enhances the network's learning ability for complex data distribution through high-order activation functions and heavy parameter strategies, thereby improving the overall robustness of the model.

[0047] In this embodiment, the present invention uses a topological self-organizing structure-based generative adversarial network for data expansion, and the topological self-organizing structure-based generative adversarial network includes: a generator (G) and a discriminator (D). The input of the generator is a random noise vector extracted from a prior distribution, and the discriminator makes a judgment based on the vectorized original training sample and the output of the generator.

[0048] like Figure 1 As shown, the data expansion method of the generative adversarial network based on the topological self-organizing structure includes steps S11-S15.

[0049] Step S11: In the initial stage, the generator and discriminator of the data augmentation model are trained in a traditional adversarial manner. The generator tries to generate samples that are sufficient to deceive the discriminator, and the discriminator strives to distinguish between real samples and generated samples. Multiple rounds of iterations are performed to gradually improve the capabilities of the generator and the discriminator.

[0050] Specifically, the initial adversarial training step is expressed as:

[0051] G(z;θ g )=xfake ,

[0052] D(x;θ d ) = Sig(x),

[0053] In the formula, G represents the generator, and z is the output from the prior distribution p z The random noise vector extracted from g is the parameter of the generator; x fake is the fake sample output by the generator; D represents the discriminator, x is the input sample, θ d are the parameters of the discriminator, and Sig() is the Sigmoid activation function.

[0054] It should be noted that the adversarial training methods of traditional generative adversarial networks include:

[0055] (1) Initialize the weights of the generator and discriminator;

[0056] (2) Training the discriminator: Extract a batch of real samples from the real data set, and then generate a batch of pseudo samples from the generator; use this batch of real samples and pseudo samples to train the discriminator respectively, so that the discriminator can distinguish between real samples and pseudo samples. This is usually achieved by minimizing the binary cross entropy loss function.

[0057] (3) Train the generator: Generate a batch of pseudo samples; use these pseudo samples and the discriminator’s feedback to update the generator’s parameters. The goal of the generator is to generate samples that can deceive the discriminator, which is usually achieved by minimizing the probability that the samples generated by the generator are correctly identified by the discriminator.

[0058] Step S12: After a preset number of iterations, the data expansion model has a certain degree of stability. Then, a nonlinear feature entanglement decoupling strategy is adopted to re-encode the features and break the linear relationship between the original features to generate more complex and varied data samples.

[0059] Among them, the nonlinear feature entanglement decoupling strategy is expressed as:

[0060] x ent =W ent ·x fake +b ent ,

[0061] x dis =ReLU(x ent ),

[0062] Where W ent and b ent are the weights and biases of the feature entanglement layer respectively; ReLU() is the linear rectification function; x ent is the intermediate feature vector after feature entanglement; xdis is the decoupled eigenvector.

[0063] Furthermore, the weight W of the feature entanglement layer ent and bias b ent The calculation method is expressed as:

[0064] W ent =UΣV T ,

[0065] b ent =μ x -W ent μ z ,

[0066] Where U, Σ and V T Obtained by performing singular value decomposition on the covariance matrix of the generated samples; μ x is the mean of the input features, μ z is the mean of the noise vector z.

[0067] Step S13: After the generated data samples are evaluated by the discriminator, the evaluation results are fed back to the generator; the generator adjusts its parameters based on the feedback to optimize the quality and diversity of the generated samples.

[0068] Specifically, the feedback regulation method is expressed as:

[0069] loss d = -logD(x real )-log(1-D(G(z))),

[0070] loss g = -logD(G(z)),

[0071]

[0072] In the formula, loss d is the loss function of the discriminator, loss g is the loss function of the generator; x real is a real sample; α d and α g are the learning rates of the discriminator and the generator respectively; and represents the gradient of the loss function with respect to the respective parameters. Preferably, α d0 and α g0 Set to 0.03 and 0.01 respectively.

[0073] Furthermore, the learning rate α d and α g It is set by dynamic adjustment, and the adjustment method is expressed as:

[0074]

[0075] In the formula, α d0 and α g0 are the initial learning rates of the discriminator and the generator respectively; β d and β g is the learning rate decay coefficient, and t is the number of training rounds.

[0076] Step S14: adopt a progressive training strategy to further improve the quality of generated data samples, gradually transitioning from the generation of simple low-dimensional features to the generation of complex high-dimensional features.

[0077] Among them, the progressive training strategy is expressed as:

[0078] x prog =λ prog ·x dis +(1-λ prog )·x real ,

[0079] In the formula, x prog is the sample generated by progressive training; prog It is a coefficient that controls the fusion ratio of real samples and generated samples.

[0080] Furthermore, the coefficient λ of the fusion ratio prog The calculation method is expressed as:

[0081]

[0082] In the formula, γ c is a parameter that controls the convergence speed of the coefficients. Preferably, γ c Set to 3.

[0083] Step S15: Repeat the above steps S12-S14 until the preset stop iteration condition is met, indicating that the data augmentation model training is completed. For example, in one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0084] Therefore, after the data expansion model training is completed, the trained data expansion model is used to increase the number of samples. For example, in one embodiment, assuming that the original collected samples are 800, and the data expansion model expands and generates 200 samples, the expanded data set contains 1000 samples.

[0085] In summary, the generative adversarial network data expansion method based on topological self-organizing structure provided by the present invention adopts a generative adversarial network based on topological self-organizing structure for data expansion, which effectively solves the problem of insufficient training samples. Through the nonlinear feature entanglement decoupling strategy, it breaks the linear correlation in the generated samples, generates samples with more complex distribution, improves the diversity and quality of the samples, and enhances the generalization ability of the model for subsequent training of the expanded samples.

[0086] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0087] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0088] Those of ordinary skill in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0089] In the embodiments provided by the present invention, it should be understood that the disclosed systems or devices / terminal equipment and methods can be implemented in other ways. For example, the system or device / terminal equipment embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the system or unit can be electrical, mechanical or other forms.

[0090] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0091] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0092] The present invention is not limited to what is described in the specification and implementation modes, and therefore additional advantages and modifications can be easily realized by those skilled in the art. Therefore, without departing from the spirit and scope of the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details, representative devices, and illustrative examples shown and described herein.

Claims

1. A data expansion method for a generative adversarial network based on a topological self-organizing structure, applied to a data expansion model based on a generative adversarial network, characterized in that: include: Step S11: In the initial stage, the generator and discriminator of the data augmentation model are trained in a traditional adversarial manner. The generator tries to generate samples that are sufficient to deceive the discriminator, and the discriminator tries to distinguish between real samples and generated samples. Multiple rounds of iterations are performed to gradually improve the capabilities of the generator and the discriminator. Step S12: After a preset number of iterations, a nonlinear feature entanglement decoupling strategy is used to re-encode the features and break the linear relationship between the original features to generate more complex and varied data samples; Step S13: After the generated data samples are evaluated by the discriminator, the evaluation results are fed back to the generator; The generator adjusts its parameters based on the feedback to optimize the quality and diversity of generated samples; Step S14: using a progressive training strategy to further improve the quality of generated data samples, transitioning from low-dimensional features to high-dimensional features; Step S15: Repeat the above steps S12-S14 until the preset stop iteration condition is met, which means that the data expansion model training is completed.

2. The method for data expansion of a generative adversarial network based on a topological self-organizing structure according to claim 1, characterized in that: In step S11, the initial adversarial training step is expressed as: G(z;θ g )=x fake , D(x;θ d )=Mr(x), In the formula, G represents the generator, and z is the output from the prior distribution p z The random noise vector extracted from g is the parameter of the generator; x fake is the fake sample output by the generator; D represents the discriminator, x is the input sample, θ d are the parameters of the discriminator, and Sig() is the Sigmoid activation function.

3. The method for data expansion of a generative adversarial network based on a topological self-organizing structure according to claim 1, characterized in that: In step S12, the nonlinear characteristic entanglement decoupling strategy is expressed as: x ent =W ent ·x fake +b ent , x dis =ReLU(x ent ), Where W ent and b ent are the weights and biases of the feature entanglement layer respectively; ReLU() is the linear rectification function; x ent is the intermediate feature vector after feature entanglement; x dis is the decoupled eigenvector.

4. The method for data expansion of a generative adversarial network based on a topological self-organizing structure according to claim 3, characterized in that: The weight W of the feature entanglement layer ent and bias b ent The calculation method is expressed as: W ent =UΣV T , b ent =μ x -W ent m z , Where U, Σ and V T Obtained by performing singular value decomposition on the covariance matrix of the generated samples; μ x is the mean of the input features, μ z is the mean of the noise vector z.

5. The method for data expansion of a generative adversarial network based on a topological self-organizing structure as claimed in claim 3, characterized in that: In step S13, the feedback adjustment method is expressed as: loss d =-logD(x real )-log(1-D(G(z))), loss g =-logD(G(z)), In the formula, loss d is the loss function of the discriminator, loss g is the loss function of the generator; x real is a real sample; α d and α g are the learning rates of the discriminator and the generator respectively; and represents the gradient of the loss function with respect to their respective parameters.

6. The method for data expansion of a generative adversarial network based on a topological self-organizing structure according to claim 5, characterized in that: Learning rate α d and α g It is set by dynamic adjustment, and the adjustment method is expressed as: In the formula, α d0 and α g0 are the initial learning rates of the discriminator and generator respectively; β d and β g is the learning rate decay coefficient, and t is the number of training rounds.

7. The method for data expansion of a generative adversarial network based on a topological self-organizing structure according to claim 6, characterized in that: α d0 and α g0 Set to 0.03 and 0.01 respectively.

8. The method for data expansion of a generative adversarial network based on a topological self-organizing structure according to claim 5, characterized in that: In step S14, the progressive training strategy is expressed as: x prog =λ prog ·x dis +(1-l prog )·x real , In the formula, x prog is the sample generated by progressive training; prog It is a coefficient that controls the fusion ratio of real samples and generated samples.

9. The method for data expansion of a generative adversarial network based on a topological self-organizing structure according to claim 8, characterized in that: Fusion ratio coefficient λ prog The calculation method is expressed as: In the formula, γ c is a parameter that controls the convergence speed of the coefficients.

10. The method for data expansion of a generative adversarial network based on a topological self-organizing structure according to claim 9, characterized in that: γ c Set to 3.

Images