Multi-source data aggregation self-coding enhancement method
By constructing a deep learning autoencoder network structure and utilizing the common features of multi-source datasets, the problem of aggregating small datasets in deep learning and machine learning is solved, achieving effective enhancement of multi-source datasets and generating more representative and diverse datasets, supporting small-sample training.
Patent Information
- Application Number
- CN202410189189.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-20
- Publication Date
- 2026-01-30
AI Technical Summary
In the fields of deep learning and machine learning, especially in the construction of material image datasets, the aggregation of datasets is complicated due to feature mismatch, partial data loss, or the use of different encoding methods in small datasets, which limits the application of deep learning.
By constructing a deep learning autoencoder network structure and utilizing the common features among multiple source datasets, the system employs autoencoders and generative adversarial networks to effectively enhance multiple source datasets. This includes identifying common features, normalization processing, autoencoder network construction, minimizing reconstruction error optimization, generating synthetic features, and combining data using hybrid mechanisms.
It successfully overcomes the limitations of small sample training data, achieves effective aggregation of multi-source datasets, and generates more representative and diverse augmented datasets, supporting deep learning and machine learning training in small sample environments.
Smart Images
Figure CN121436047A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning, and specifically to a method for enhancing multi-source data aggregation autoencoders. Background Technology
[0002] In the field of machine learning, especially deep learning, obtaining sufficient training samples is crucial when processing data. In practical applications, the size of the training samples directly impacts the performance of machine learning and deep learning models. Taking the collection of materials image data as an example, the complex and costly process of constructing large datasets has significantly limited the application of deep learning in materials research.
[0003] To address this issue, aggregating multiple existing small datasets to construct large datasets is a feasible solution. However, dataset aggregation becomes extremely complex when the features of the various small datasets do not match, some data is missing, or different encoding methods are used. For datasets within the same technical field, there are often some shared features, making it possible to construct large datasets by aggregating small datasets.
[0004] Taking an image dataset studying the thermal conductivity properties of thermal barrier coating materials as an example, it contains information such as grain boundary density, porosity, and aspect ratio. Even in datasets studying the coefficient of thermal expansion of materials, the data may not be completely identical, but they may still contain common features such as grain boundary density and porosity, leading to feature overlap.
[0005] In addition to these common features, each dataset may also possess unique characteristics, complicating direct aggregation. However, the shared features among these datasets may provide sufficient information. Based on these shared features, a deep learning model can be trained to reconstruct the original dataset, thereby achieving effective aggregation of different datasets. Summary of the Invention
[0006] The purpose of this invention is to address the data augmentation challenge in deep learning and machine learning by providing a multi-source data aggregation autoencoder augmentation method. This method, starting from the perspective of integrating multiple data sources, constructs a deep learning autoencoder network structure and utilizes the common features among the multi-source datasets as input, successfully achieving effective augmentation of multi-source datasets. This invention shows broad application prospects in the field of deep learning and machine learning with small sample sizes.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] To address the challenges of data augmentation in deep learning and machine learning, this invention provides a multi-source data aggregation autoencoder augmentation method. This method, starting from the perspective of integrating multiple data sources, constructs a deep learning autoencoder network structure and utilizes the common features among the multi-source datasets as input, successfully achieving effective augmentation of multi-source datasets. This invention shows broad application prospects in the field of small-sample deep learning and machine learning.
[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:
[0010] A method for enhancing multi-source data aggregation using autoencoders includes the following steps:
[0011] S1: Identify the original common features of multi-source datasets and obtain the original common feature data;
[0012] S2: Normalize the original common feature data to generate common feature data;
[0013] S3: Construct an autoencoder network using an autoencoder and a decoder;
[0014] S4: Input the common feature data into the autoencoder network to complete the compression and reconstruction of the common feature data;
[0015] S5: Randomly select common feature data from a dataset after compression and reconstruction, and optimize the autoencoder network using a loss function that minimizes the reconstruction error;
[0016] S6: After the autoencoder network is optimized, the common feature data of the remaining dataset after compression and reconstruction is used to generate synthetic feature data through the discriminator of Generative Adversarial Network (GAN).
[0017] S7: Introduces a hybrid mechanism to linearly or nonlinearly combine synthetic feature data and original common feature data to construct an enhanced dataset.
[0018] In a preferred embodiment, the normalization process of the original common feature data in step S2 specifically includes normalizing the original common feature data to the range of 0 to 1.
[0019] In a preferred embodiment, step S3 specifically includes:
[0020] S31: The autoencoder includes a convolutional layer, a pooling layer, and an activation function;
[0021] The autoencoder described in step S31 is used for compressing common feature data; the autoencoder network is represented by the symbol E. This represents an autoencoder function with parameter θ1.
[0022] S32: The decoder includes a fully connected layer, a deconvolutional layer, and a convolutional layer for output;
[0023] The decoder described in step S32 is used for reconstructing common feature data; the decoder network is represented by the symbol D. This represents a decoder function with parameter θ2.
[0024] In a preferred embodiment, step S4 specifically includes:
[0025] S41: Input the common feature data X into the autoencoder and convert it into a latent space representation Z, i.e.
[0026] S42: Input the latent space representation Z into the decoder and convert it into reconstructed data X′, i.e.
[0027] In a preferred embodiment, step S5 specifically includes:
[0028] S51: Randomly select common feature data Y from a dataset. normalized ;
[0029] S52: Transfer the common feature data Y normalized Input autoencoder network;
[0030] S53: Minimize the reconstruction error loss function using a stochastic gradient descent optimizer, and update the parameters θ1 and θ2 of the autoencoder network.
[0031] In a preferred embodiment, step S53, which involves minimizing the reconstruction error loss function using a stochastic gradient descent optimizer, specifically includes:
[0032] S53a: The Mean Squared Error (MSE) function is used to measure the reconstruction error, as shown in the following formula:
[0033]
[0034] Among them, L recon Let X represent the reconstruction error loss function. i This represents the i-th sample in the input dataset, and N represents the number of samples.
[0035] S53b: The stochastic gradient descent optimizer optimizes the objective as follows:
[0036]
[0037] In a preferred embodiment, step S6 specifically includes:
[0038] S61: The generative adversarial network discriminator includes a generator network G and a discriminator network D;
[0039] S62: The generator network G receives the latent spatial variables Z of the remaining common feature dataset as input and outputs synthetic feature data Z′;
[0040] S63: The discriminator network D evaluates the authenticity of the synthesized feature data Z′.
[0041] In a preferred embodiment, the discriminator network D in step S63 includes convolutional layers and fully connected layers for binary classification.
[0042] In a preferred embodiment, the mixing mechanism in step S7
[0043] Its linear combination representation is as follows:
[0044] X enhanced =(1-α)·X+α·Z'
[0045] Its nonlinear combination representation is as follows:
[0046] X enhanced = (1-α)·sin(X)+α·Z'
[0047] Among them, X enhanced Let Z represent the augmented dataset, α represent the mixing parameters, X represent the common feature data, and Z′ represent the synthetic feature data.
[0048] Compared with the prior art, the present invention has the following obvious and prominent substantive features and significant advantages:
[0049] This invention proposes an innovative method to address the challenge of data augmentation for few-sample training data in deep learning and machine learning. First, by constructing an autoencoder structure for a deep learning neural network, it reconstructs multi-source datasets based on their original common feature data, starting from the perspective of integrating multiple data sources. Then, the trained neural network aggregates the various datasets. Specifically, a convolutional autoencoder structure, including encoder and decoder parts, is used for efficient data compression and reconstruction. Subsequently, during the network training phase, the autoencoder network is optimized by minimizing the reconstruction error loss function to restore the original input data to the greatest extent possible. In the data augmentation phase, a generative adversarial network (GAN) discriminator is introduced to generate more realistic synthetic feature data. The generator receives latent spatial variables and outputs synthetic feature data, while the discriminator evaluates the realism of the synthetic features. Finally, an augmented dataset is constructed through a hybrid mechanism, linearly or non-linearly combining the synthetic feature data and common feature data to obtain more representative and diverse data representations. This method successfully overcomes the limitations of few-sample training data for deep learning and machine learning, providing an innovative solution for effective training in small-sample environments. Attached Figure Description
[0050] The accompanying drawings, which constitute a part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0051] Figure 1 This is a flowchart of a preferred embodiment of the multi-source data aggregation self-encoding enhancement method of the present invention. Detailed Implementation
[0052] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be reviewed and described completely below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely a part of the implementation examples of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the protection scope of the present invention.
[0053] A method for enhancing multi-source data aggregation using autoencoders includes the following steps:
[0054] S1: Identify the original common features of multi-source datasets and obtain the original common feature data;
[0055] S2: Normalize the original common feature data to generate common feature data;
[0056] S3: Construct an autoencoder network using an autoencoder and a decoder;
[0057] S4: Input the common feature data into the autoencoder network to complete the compression and reconstruction of the common feature data;
[0058] S5: Randomly select common feature data from a dataset after compression and reconstruction, and optimize the autoencoder network using a loss function that minimizes the reconstruction error;
[0059] S6: After the autoencoder network is optimized, the common feature data of the remaining dataset after compression and reconstruction is used to assist in the generation of synthetic feature data through a generative adversarial network discriminator;
[0060] S7: Introduces a hybrid mechanism to linearly or nonlinearly combine synthetic feature data and common feature data to construct an enhanced dataset.
[0061] In this embodiment, step S1 identifies the original common features of the multi-source datasets and obtains the original common feature data. The multi-source datasets are multiple datasets within the same field, such as datasets studying the thermal expansion coefficient of thermal barrier coatings, datasets studying the thermal conductivity of thermal barrier coating materials, and datasets studying the strength properties of copper-aluminum alloy materials in the field of materials science. For datasets studying thermal barrier coatings, common features include grain boundary density, porosity, and aspect ratio, etc. For datasets studying copper-aluminum alloys, common features include the physical properties and chemical composition of copper-aluminum alloy materials, etc. The obtained original common features are the overlapping feature portions among the collected multi-source datasets within the same field.
[0062] In this embodiment, the normalization process of the original common feature data in step S2 specifically includes normalizing the original common feature data to the range of 0 to 1, and the data after normalization is the feature data.
[0063] In this embodiment, step S3, which constructs an autoencoder network using an autoencoder and a decoder, specifically includes:
[0064] S31: The autoencoder includes a convolutional layer, a pooling layer, and an activation function;
[0065] The autoencoder described in step S31 is used for compressing common feature data; the autoencoder network is represented by the symbol E. This represents an autoencoder function with parameter θ1;
[0066] The autoencoder structure involves components such as convolutional layers, pooling layers, and activation functions, which work together to map input data to a latent space. The autoencoder has three convolutional layers, which perform convolution operations on the data using an average pooling strategy and a Leaky ReLU activation function. The kernel size of each convolutional layer in the autoencoder is set to 4×4, the stride to 2, and the number of channels to 64, 128, and 256 respectively. The slope of the Leaky ReLU activation function is 0.2.
[0067] S32: The decoder includes a deconvolutional layer and an activation function;
[0068] The decoder described in step S32 is used for reconstructing common feature data; the decoder network is represented by the symbol D. This represents a decoder function with parameter θ2.
[0069] In this embodiment, the decoder involves components such as fully connected layers, deconvolutional layers, and convolutional layers. There is one convolutional layer for data output; three deconvolutional layers in the decoder are used to reconstruct the latent vectors; the kernel size of the convolutional layers in the decoder is 4×4. The decoder structure is designed to map the representation of the latent space back to the original data space, thereby facilitating the generation of subsequent synthetic data. After processing the latent vectors through the fully connected layers, the data is then passed to the deconvolutional layers for further processing, and the data reconstruction is completed through the convolutional layers.
[0070] Page 7 of the detailed implementation section of the disclosure document describes the decoder as employing four components: a fully connected layer, a reshape layer, three deconvolutional layers, and a convolutional layer for output. This is inconsistent with the description in the invention content section that the decoder consists of a fully connected layer, a deconvolutional layer, and a convolutional layer.
[0071] In this embodiment, step S4 inputs the common feature data into the autoencoder network to complete the compression and reconstruction of the common feature data, specifically including:
[0072] S41: Input the common feature data X into the autoencoder and convert it into a latent space representation Z, i.e.
[0073] S42: Input the latent space representation Z into the decoder and convert it into reconstructed data X′, i.e.
[0074] In this embodiment, common feature data is used as input, and the input format is 224×224×n, where n is the number of channels. In this embodiment, the number of channels is 3.
[0075] In this embodiment, step S5 randomly selects common feature data after compression and reconstruction of a dataset, and optimizes the autoencoder network using a loss function that minimizes the reconstruction error. Specifically, this includes:
[0076] S51: Randomly select a dataset Y normalized Common characteristic data;
[0077] S52: Transfer the common feature data Y normalized Input autoencoder network;
[0078] S53: Minimize the reconstruction error loss function using a stochastic gradient descent optimizer, and update the parameters θ1 and θ2 of the autoencoder network.
[0079] In this embodiment, step S53 specifically includes:
[0080] S53a: The Mean Squared Error (MSE) function is used to measure the reconstruction error, as shown in the following formula:
[0081]
[0082] Among them, L recon Let X represent the reconstruction error loss function. i This represents the i-th sample in the input dataset, and N represents the number of samples.
[0083] S53b: The stochastic gradient descent optimizer optimizes the objective as follows:
[0084]
[0085] In this embodiment, after the autoencoder network optimization is completed in step S6, the common feature data of the remaining dataset after compression and reconstruction is used to assist in generating synthetic feature data through a generative adversarial network discriminator, specifically including:
[0086] S61: The generative adversarial network discriminator includes a generator network G and a discriminator network D;
[0087] S62: The generator network G receives the latent spatial variables Z of the remaining common feature dataset as input and outputs synthetic feature data Z′;
[0088] S63: The discriminator network D evaluates the authenticity of the synthesized feature data Z′.
[0089] In this embodiment, the discriminator network D in step S63 includes three convolutional layers and one fully connected layer for binary classification, enabling the discriminator network to effectively evaluate the authenticity of the input data. The convolutional kernel size of the convolutional layers in the discriminator network D is 4×4, the stride is 2, and the channels are 64, 128, and 256 respectively. In step S6, the discriminator judges the generated synthetic features. Only when it cannot distinguish between the synthetic feature dataset and the original common feature dataset will the discriminator determine that the synthetic feature data can be used as qualified generated feature data for subsequent data augmentation.
[0090] In this embodiment, step S7 introduces a hybrid mechanism to linearly or nonlinearly combine synthetic feature data and common feature data to construct an enhanced dataset. Specifically, the hybrid mechanism is as follows:
[0091] Its linear combination representation is as follows:
[0092] X enhanced = (1-α)·X+α·Z′
[0093] Its nonlinear combination representation is as follows:
[0094] X enhanced = (1-α)·sin(X)+α·Z′
[0095] Among them, X enhanced Let Z represent the augmented dataset, α represent the mixing parameters, X represent the common feature data, and Z′ represent the synthetic feature data.
[0096] This embodiment is implemented based on the PyTorch framework and the VSCode development environment.
[0097] This invention addresses the scarcity of small-sample training data in machine learning and deep learning by proposing a multi-source data aggregation autoencoder augmentation method. This method aggregates multiple small datasets from the same domain using autoencoder technology, aiming to optimize training results for deep learning and machine learning using small-sample datasets. First, an autoencoder network is constructed to reconstruct multi-source datasets based on common features from multiple data sources. Then, a trained model is used to aggregate these datasets. Specifically, an autoencoder network, including an encoder and a decoder, is used for efficient data compression and reconstruction, respectively. Next, during network training, the autoencoder network model is optimized by minimizing the reconstruction error loss function to maximize the restoration of the original input data. In the data augmentation stage, a generative adversarial network (GAN) discriminator is introduced to generate more realistic synthetic feature data, which is then evaluated by the discriminator. Finally, a hybrid mechanism is introduced to augment the dataset, combining the evaluated synthetic feature data and the original feature data in a linear or non-linear manner to obtain more representative and diverse data representations. This method successfully overcomes the limitations of small sample training data on deep learning and machine learning, providing an innovative solution for effective training in small sample environments.
[0098] The specific embodiments of the present invention have been described in detail above, but they are only examples, and the present invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to this utility model are also within the scope of the present invention. Therefore, all equivalent changes and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.
Claims
1. A multi-source data aggregation auto-encoding enhancement method, characterized in that, The method comprises the following steps: S1: identifying original common characteristics of multi-source data sets, and obtaining original common characteristic data; S2: normalizing the original common characteristic data to generate common characteristic data; S3: constructing a self-encoder network through a self-encoder and a decoder; S4: inputting the common characteristic data into the self-encoder network to complete compression and reconstruction of the common characteristic data; S5: randomly selecting common characteristic data of a data set after compression and reconstruction, and optimizing the self-encoder network by using a loss function of minimizing reconstruction error; S6: after optimization of the self-encoder network, the common characteristic data after compression and reconstruction is used to assist in generating synthetic characteristic data through a discriminator of a generative adversarial network; S7: introducing a mixing mechanism to linearly or nonlinearly combine the synthetic characteristic data and the common characteristic data to construct an enhanced data set.
2. The multi-source data aggregated self-encoding enhancement method of claim 1, wherein, The normalization of the original common characteristic data in step S2 comprises normalizing the original common characteristic data to a range of 0 to 1.
3. The multi-source data aggregated self-encoding enhancement method of claim 1, wherein, The self-encoder network in step S3 comprises: S31: the self-encoder is composed of a convolutional layer, a pooling layer and an activation function; S32: the decoder is composed of a deconvolutional layer and an activation function.
4. The multi-source data aggregated self-encoding enhancement method of claim 3, wherein, The autoencoder described in step S31 is used for compression of the common feature data; the autoencoder network is denoted by the symbol E, denotes the autoencoder function with the parameter θ1; The decoder described in step S32 is used for reconstruction of the common feature data; let D denote the decoder network, denotes the decoder function with parameter 02.
5. The multi-source data aggregation auto-encoding enhancement method of claim 4, wherein, In step S4, the common characteristic data is input into the self-encoder network. S41: input the common feature data X into the autoencoder, and convert it into a latent space representation Z, i.e. S42: The latent space representation Z is input into the decoder, which converts it into the reconstructed data X', i.e.
6. The multi-source data aggregation auto-encoding enhancement method of claim 4, wherein, In step S5, the common characteristic data of a data set is randomly selected. S52: inputting the common characteristic data into the self-encoder network; S53: minimizing the reconstruction error loss function by using a stochastic gradient descent optimizer to update the parameters θ1 and θ2 of the self-encoder network. In step S53, the following steps are included:
7. The multi-source data aggregation auto-encoding enhancement method of claim 7, wherein, S53a: a square loss function (Mean Squared Error) is used to measure the reconstruction error, and the formula is as follows: S53b: the optimization target of the stochastic gradient descent optimizer is as follows: wherein L recon represents the reconstruction error loss function, X i represents the i-th sample in the input dataset, and N represents the number of samples; In step S6, the following steps are included:
8. The multi-source data aggregation auto-encoding enhancement method of claim 1, wherein, S61: the generative adversarial network discriminator comprises a generator network G and a discriminator network D; S62: the generator network G receives a latent space variable Z of a remaining common characteristic data set as input and outputs synthetic characteristic data Z'; S63: the discriminator network D evaluates the authenticity of the synthetic characteristic data Z'. In step S63, the discriminator network D is composed of a convolutional layer and a fully connected layer for classification.
9. The multi-source data aggregation auto-encoding enhancement method of claim 8, wherein, In step S7, the mixing mechanism is linear combination or nonlinear combination.
10. The multi-source data aggregation auto-encoding enhancement method of claim 9, wherein, The linear combination is represented as follows: The nonlinear combination is represented as follows: X enhanced = (1 - a) · X + a · Z' X enhanced = (1 - a) - sin(X) + a - Z' where X enhanced represents the enhanced dataset, a represents the mixing parameter, X represents the original features, and Z' represents the generated synthetic feature data.