Layout generation method and device based on recursive self-encoding, equipment and storage medium

Abstract

The invention relates to a layout generation method based on recursive self-encoding, which comprises the following steps: acquiring equipment images and equipment state data of underwater target equipment, preprocessing the equipment images, extracting damage area images of the underwater target equipment from the equipment images, extracting damage features from the damage area images and the equipment state data, and generating layout features of the underwater target equipment; the damage features comprise the shape, color and texture features of the damage area, analyzing the damage features by using a pre-trained deep learning model, generating a damage evaluation result, and determining the damage level of the underwater target equipment. According to the invention, intelligent detection and rapid evaluation of underwater equipment damage can be realized, the accuracy and efficiency of damage detection are improved, the subjectivity and error of manual operation are reduced, the timeliness and reliability of equipment state monitoring are ensured, and a scientific basis is provided for maintenance and insurance claim settlement of underwater equipment.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a layout generation method, device, equipment and storage medium based on recursive autoencoding. Background Art

[0002] In the financial industry, with the rapid growth of data volume, financial institutions are faced with the huge challenge of how to efficiently process and analyze multi-source data to provide accurate decision support. In the existing technology, although a large number of data processing and analysis methods have been applied, such as machine learning models and deep learning algorithms, these methods still have the following major problems when facing the complexity and high-dimensional characteristics of financial data:

[0003] Insufficient data privacy protection: Traditional data analysis methods usually require data from multiple sources to be centralized on a single platform for processing, which has brought serious privacy and security issues in the financial sector. In particular, the security of data sharing between different financial institutions is difficult to guarantee, which reduces the willingness to cooperate on data and thus affects the training effect of the model.

[0004] Limited model generalization ability: Existing machine learning models often lack the ability to generalize when faced with the ever-changing financial market. Especially when dealing with cross-institutional and cross-domain data, the model is prone to overfitting or poor performance on new data, which limits its effectiveness in practical applications.

[0005] High consumption of computing resources: Existing deep learning models, especially those trained based on large-scale data sets, often consume a lot of computing resources and time. This makes it difficult to meet the requirements of efficiency and speed in the financial field, especially in real-time data processing and decision support scenarios.

[0006] Inefficient data processing: Financial data is highly heterogeneous and complex. Existing models often require a lot of preprocessing and feature engineering when processing this data, which reduces the overall data processing efficiency. In addition, effective feature extraction from high-dimensional data is still a challenge, and existing methods are prone to losing key information when extracting important features.

[0007] In summary, although current technologies have been applied to the processing and analysis of financial data to a certain extent, there are still obvious shortcomings in data privacy protection, model generalization capabilities, computing resource consumption, and data processing efficiency. Summary of the invention

[0008] The main purpose of the present invention is to provide a layout generation method, device, equipment and storage medium based on recursive autoencoding, aiming to solve the technical problem that the existing technology is difficult to perform effective model training and generate layout data while protecting the privacy of multiple data holders in financial data processing.

[0009] To achieve the above object, the present invention provides a layout generation method based on recursive autoencoding, comprising:

[0010] Acquire local layout data from multiple training data holders respectively, where the local layout data includes location information, size information, and type information of a bounding box;

[0011] Based on each local layout data, a local recursive autoencoder of each training data holder is trained to obtain a local model parameter of each local recursive autoencoder;

[0012] Upload all local model parameters to the federal server, aggregate the local model parameters through the federal server, and generate global model parameters;

[0013] Sending the global model parameters to each training data holder through the federal server, training a local recursive autoencoder based on the global model parameters, and updating the local model parameters of each local recursive autoencoder;

[0014] The training, uploading, aggregating and updating operations are repeatedly performed until a predetermined stopping condition is met, the training process is stopped, a target local recursive autoencoder of each training data holder is obtained, and layout data is obtained based on the target local recursive autoencoder.

[0015] In one embodiment, the local recursive autoencoder of each training data holder is trained based on each local layout data to obtain the local model parameters of each local recursive autoencoder, including:

[0016] Use a local recursive autoencoder to extract features from the local layout data of each training data holder to obtain a low-dimensional feature vector;

[0017] Reconstructing the low-dimensional feature vector through a local recursive autoencoder to generate reconstructed layout data for each local layout data;

[0018] The reconstruction error and KL divergence are calculated based on the local layout data and the corresponding reconstructed layout data, and each local recursive autoencoder is trained to obtain the local model parameters of each local recursive autoencoder.

[0019] In one embodiment, a local recursive autoencoder is used to extract features from the local layout data of each training data holder to obtain a low-dimensional feature vector, including:

[0020] Input the local layout data of each training data holder into the local recursive autoencoder and generate the mean and standard deviation representing the distribution of the local layout data in the latent space;

[0021] Generate a random variable that satisfies a standard normal distribution, and multiply the random variable by the standard deviation to obtain a perturbation distribution value for adjusting feature representation;

[0022] The disturbance distribution value is added to the mean value to generate the low-dimensional feature vector, which contains key feature information of the local layout data.

[0023] In one embodiment, the low-dimensional feature vector is reconstructed by a local recursive autoencoder to generate reconstructed layout data for each local layout data, including:

[0024] Inputting the low-dimensional feature vector into the local recursive autoencoder, and performing a deconvolution operation on the low-dimensional feature vector using a preset convolution kernel and stride parameter to generate a feature map with a higher resolution;

[0025] Apply activation function to perform nonlinear transformation on the generated feature map;

[0026] Repeat the deconvolution operation to increase the resolution of the feature map layer by layer. After each deconvolution operation, continue to apply the activation function for nonlinear transformation until the size of the feature map is completely restored to the same size as the original local layout data;

[0027] The final feature map generated after multiple deconvolution operations is input into a linear activation function and converted into the reconstructed layout data.

[0028] In one embodiment, the reconstruction error and KL divergence are calculated based on the local layout data and the corresponding reconstructed layout data, and each local recursive autoencoder is trained, including:

[0029] Compare each data point in the local layout data with the corresponding data point in the reconstructed layout data, and calculate the difference between each pair of data points;

[0030] The difference between each data point is squared, and the squared differences of all data points are added together and divided by the total number of data points to obtain an average square error, which is used as the reconstruction error;

[0031] Extract the distribution parameters in the latent space, including the mean and standard deviation, from the encoder output of the local recursive autoencoder, and calculate the difference between the distribution parameters and the standard normal distribution to obtain the KL divergence;

[0032] Performing a weighted summation of the reconstruction error and the KL divergence to form a total loss function;

[0033] Using the gradient descent algorithm, the gradient of the total loss function with respect to the model parameters is calculated, and the model parameters of each local recursive autoencoder are adjusted by the gradient and the set learning rate.

[0034] In one embodiment, the local model parameters are aggregated by the federal server to generate global model parameters, including:

[0035] Calculate the weight of each local recursive autoencoder according to the ratio of the amount of local layout data of each training data holder to the total amount of local layout data of all training data holders;

[0036] Based on the weight of each local recursive autoencoder, the local model parameters of each local recursive autoencoder are multiplied by the corresponding weight through the federal server, and all weighted model parameters are summed to finally generate the global model parameters.

[0037] In one embodiment, training the local recursive autoencoders based on the global model parameters and updating the local model parameters of each local recursive autoencoder includes:

[0038] Using the global model parameters as initial model parameters of the local recursive autoencoder, and combining with the local layout data, calculating the loss function of the initial model parameters on the local layout data;

[0039] Calculate the loss function equivalent to the gradient of the initial model parameters through the gradient descent algorithm;

[0040] Setting a learning rate for the local recursive autoencoder, multiplying the gradient by the learning rate to generate an update value for adjusting the initial model parameters;

[0041] The initial model parameters are adjusted based on the updated values, and the local model parameters of each local recursive autoencoder are updated.

[0042] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a layout generation device based on recursive self-encoding, wherein the layout generation device based on recursive self-encoding comprises a memory, a processor, and a layout generation program based on recursive self-encoding stored in the memory and executable on the processor, wherein the layout generation program based on recursive self-encoding implements the steps of the layout generation method based on recursive self-encoding as described above when the layout generation program based on recursive self-encoding is executed by the processor.

[0043] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer storage medium, on which a layout generation program based on recursive self-encoding is stored, and when the layout generation program based on recursive self-encoding is executed by a processor, the steps of the layout generation method based on recursive self-encoding as described above are implemented.

[0044] Beneficial effects: The present invention relates to a layout generation method based on recursive autoencoder, which obtains local model parameters by obtaining local layout data from multiple training data holders and using local recursive autoencoders to train the layout data. The local model parameters of each holder are uploaded to the federal server, and global model parameters are generated through aggregation, and the global model parameters are sent to each holder to update the local model. The training, uploading, aggregation and updating operations are repeated until the stop condition is reached, and the target recursive autoencoder of each holder is obtained, and the final layout data is obtained based on the autoencoder. The present invention effectively solves the problem of achieving efficient generation of layout data across institutions while protecting data privacy, and improves the generalization ability and application performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0046] Figure 1 It is a flow chart of an embodiment of a layout generation method based on recursive autoencoding of the present invention;

[0047] Figure 2 It is a functional module diagram of a preferred embodiment of a layout generation device based on recursive autoencoding of the present invention;

[0048] Figure 3 It is a structural schematic diagram of the device hardware operating environment involved in the embodiment of the layout generation device based on recursive self-encoding of the present invention. DETAILED DESCRIPTION

[0049] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0050] In the financial industry, with the rapid growth of data volume, financial institutions are faced with the huge challenge of how to efficiently process and analyze multi-source data to provide accurate decision support. In the existing technology, although a large number of data processing and analysis methods have been applied, such as machine learning models and deep learning algorithms, these methods still have the following major problems when facing the complexity and high-dimensional characteristics of financial data:

[0051] Insufficient data privacy protection: Traditional data analysis methods usually require data from multiple sources to be centralized on a single platform for processing, which has brought serious privacy and security issues in the financial sector. In particular, the security of data sharing between different financial institutions is difficult to guarantee, which reduces the willingness to cooperate on data and thus affects the training effect of the model.

[0052] Limited model generalization ability: Existing machine learning models often lack the ability to generalize when faced with the ever-changing financial market. Especially when dealing with cross-institutional and cross-domain data, the model is prone to overfitting or poor performance on new data, which limits its effectiveness in practical applications.

[0053] High consumption of computing resources: Existing deep learning models, especially those trained based on large-scale data sets, often consume a lot of computing resources and time. This makes it difficult to meet the requirements of efficiency and speed in the financial field, especially in real-time data processing and decision support scenarios.

[0054] Inefficient data processing: Financial data is highly heterogeneous and complex. Existing models often require a lot of preprocessing and feature engineering when processing this data, which reduces the overall data processing efficiency. In addition, effective feature extraction from high-dimensional data is still a challenge, and existing methods are prone to losing key information when extracting important features.

[0055] In summary, although current technologies have been applied to the processing and analysis of financial data to a certain extent, there are still obvious shortcomings in data privacy protection, model generalization capabilities, computing resource consumption, and data processing efficiency.

[0056] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of a layout generation method based on recursive autoencoding provided by the present invention. It should be noted that although a logical sequence is shown in the flow chart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0057] like Figure 1 As shown, the layout generation method based on recursive autoencoding proposed by the present invention includes the following steps:

[0058] S10, respectively obtaining local layout data from a plurality of training data holders, wherein the local layout data includes position information, size information, and type information of a bounding box;

[0059] In this embodiment, each training data holder independently obtains the local layout data it has. These data mainly include the location information, size information and type information of the bounding box. A bounding box is a rectangular box used to mark a specific area in an image or layout, which defines the spatial location (location information), size (size information) and content category (type information) of the area. Specifically, it can be:

[0060] Position information: The position information of the bounding box is usually represented by the coordinates of its upper left corner (such as (x, y)), which indicate the specific position of the bounding box in the entire layout.

[0061] Size information: The size information of the bounding box includes its width and height, usually represented by "w" and "h" respectively. These data help define the spatial coverage of the bounding box in the layout, ensuring that it can accurately capture the contained content.

[0062] Type information: The type information of the bounding box identifies the content category of the area, usually represented by type, such as text, image, button, etc. In financial application scenarios, this may include text areas for transaction records, chart areas, or user information display areas. Through type information, the model can distinguish and process different categories of data content.

[0063] The collection and accurate description of these layout data are the basis for the subsequent recursive autoencoder training, ensuring that the model can extract effective features from the structured layout information for optimizing data processing and analysis results. In financial scenarios, bounding boxes can represent things like account transaction frequency, geographic location activity range, etc., helping to extract key features for further analysis.

[0064] In one specific implementation, when obtaining local layout data, the financial institution not only obtains the location information, size information, and type information of the bounding box, but also introduces time series features to capture the dynamic characteristics of data changing over time. The location information of the bounding box is represented by the geographic coordinates where the transaction or behavior occurs, the size information is the transaction amount or quantity, and the type information identifies the category of the transaction. By adding time series features, the institution is able to capture time-dependent information during model training, which enhances the performance of the model in dealing with periodic or sudden financial events. The training process of the recursive autoencoder therefore not only focuses on static data features, but also models dynamic time series data to generate richer local model parameters. After these parameters are uploaded to the federal server, the global model parameters generated by aggregation also include information on the time dimension, further enhancing the model's predictive ability.

[0065] In other specific embodiments, in the scenario of insurance claims, the system combines multidimensional space analysis technology while obtaining bounding box layout data. Location information includes the geographical location of the claim event, while size information combines the amount of loss and the scope of the event. Type information is also extended to include event categories and claims methods. Multidimensional space analysis technology helps the system to perform multi-dimensional feature fusion on the data when training the recursive autoencoder, thereby extracting more expressive feature vectors. The trained local model parameters are uploaded to the federal server, and the global model generated by aggregation can comprehensively analyze claims data across institutions and regions, providing more accurate risk assessment and claims prediction.

[0066] By obtaining the location, size, and type information of the bounding box from multiple training data holders, we can effectively capture the key features of financial data and ensure the accuracy of subsequent model training. We train these layout data through recursive autoencoders and use the federated learning architecture to aggregate and optimize model parameters, achieving efficient joint modeling under the premise of data privacy protection.

[0067] S20, training the local recursive autoencoder of each training data holder based on each local layout data to obtain local model parameters of each local recursive autoencoder;

[0068] In this embodiment, the training data holder uses its local layout data to train the recursive autoencoder model. The recursive autoencoder is a deep learning model whose structure consists of an encoder and a decoder, which can compress high-dimensional input data into low-dimensional representations and reconstruct the original data from these low-dimensional representations. The training process minimizes the difference between the input data and the reconstructed data by continuously adjusting the model parameters (such as weights and biases). Through this training, the recursive autoencoder can capture the key features of the layout data and generate corresponding local model parameters. These parameters reflect the model's learning results for the local data, and because the data is not shared with other holders, data privacy is guaranteed.

[0069] In one specific implementation, in a smart investment advisory platform, financial institutions extract layout information from clients’ portfolio data, including asset categories, investment amounts, and holding periods. During training, the recursive autoencoder not only extracts traditional low-dimensional features, but also takes into account the correlation between different asset categories and the changing trend of holding periods. By capturing the correlation between assets, the model generates local model parameters that can more accurately reflect the risk and return characteristics of the investment portfolio, ultimately providing clients with more personalized and risk-balanced investment advice.

[0070] By using local layout data to train the recursive autoencoder, it is possible to effectively extract and compress key features in high-dimensional data and generate model parameters that reflect local data features. This not only enhances the model's ability to understand data, but also ensures data privacy without sharing the original data.

[0071] S30, uploading all local model parameters to a federal server, aggregating the local model parameters through the federal server to generate global model parameters;

[0072] In this embodiment, each training data holder independently performs model training on its local data to generate local model parameters. In order to achieve multi-party joint modeling while ensuring data privacy, this embodiment uses a federated learning architecture to achieve parameter aggregation.

[0073] After completing the training of the local recursive autoencoder, each training data holder uploads the generated local model parameters to the federated server. These parameters usually include the weights and biases of the model, reflecting the understanding of the local data characteristics of each model.

[0074] After receiving the local model parameters from all holders, the federated server performs parameter aggregation. The aggregation method can be weighted average, simple average or other algorithms, depending on the amount and importance of data from each holder. The goal of aggregation is to generate a global model parameter that can integrate the model characteristics of each holder and improve the overall performance of the model.

[0075] The global model parameters generated by parameter aggregation represent the collective data characteristics of all training data holders. This global model parameter is more generalized than the local model of a single holder, can be applied to a wider range of data, and has higher prediction and analysis capabilities.

[0076] During the entire upload and aggregation process, each holder only uploads model parameters instead of original data, so data privacy is effectively protected. This mechanism ensures that the data of each holder will not be leaked or directly accessed by other holders.

[0077] In another specific implementation, in an anti-fraud detection system in which multiple insurance companies cooperate, each company uploads the local model parameters to the federal server after completing the training of the local recursive autoencoder model. In order to better cope with the differences in data characteristics among different companies, the server introduces a data similarity metric when aggregating parameters. By calculating the similarity of the data characteristics of each company, the server gives a higher weight to similar data when aggregating parameters, so that the generated global model parameters can more accurately reflect the common characteristics among different companies. Doing so helps improve the applicability and detection accuracy of the global anti-fraud model in a cross-company data environment.

[0078] In other specific embodiments, in the global smart investment advisory platform, financial institutions in various regions upload model parameters to the federal server after completing the recursive autoencoder training of local customer investment data. In order to improve the diversity and adaptability of the global model, the server adopts a hierarchical aggregation algorithm when aggregating parameters. The algorithm first groups the parameters of each institution according to regional characteristics and performs local aggregation to generate intermediate model parameters for each region. Subsequently, the server performs a second round of aggregation on these intermediate model parameters to generate a global model parameter with global applicability. Such a hierarchical aggregation method can better capture the investment behavior characteristics of different regions, and ultimately send them to financial institutions in various regions to provide customers with more diversified and personalized investment advice.

[0079] By uploading local model parameters to the federated server and aggregating them, the goal of improving the generalization ability of the model while protecting data privacy is achieved. The generated global model parameters can integrate the characteristics of multiple data parties to provide more accurate and robust model support for scenarios such as financial risk control, anti-fraud detection, and smart investment advisors. It not only effectively improves the overall performance of the model, but also ensures that sensitive information will not be leaked in data cooperation between multiple holders, thereby enhancing data security.

[0080] S40, sending the global model parameters to each training data holder through the federal server, training a local recursive autoencoder based on the global model parameters, and updating a local model parameter of each local recursive autoencoder;

[0081] In this embodiment, after completing the aggregation of the global model parameters, the federated server distributes the global model parameters back to each training data holder. After receiving the global model parameters, each holder uses them as initial model parameters or weights and continues to train and optimize the local recursive autoencoder in combination with local data.

[0082] The federated server sends the aggregated global model parameters to each training data holder through a secure transmission protocol. The global model parameters represent the comprehensive characteristics of multi-party data and have higher generalization capabilities.

[0083] Each owner uses the global model parameters as the initial weights or parameters of its local recursive autoencoder and retrains it with its own local data. This training step aims to make the model better fit the owner's specific data distribution through further optimization of local data.

[0084] After training, the model parameters of the recursive autoencoder will be updated to reflect the learning results on the local data after combining the global model parameters. These updated parameters can more accurately represent the local data characteristics while retaining the broad adaptability of the global model.

[0085] In other specific implementations, in the global smart investment advisory platform, after receiving the global investment analysis model parameters, each regional financial institution conducts multi-level training in combination with local market characteristics. The multi-level training method is divided into global parameter adjustment at the basic level and local optimization training at the advanced level. The basic level adjusts the global model parameters to adapt them to the general characteristics of the regional market; while the advanced level further uses the investment preferences and market dynamics of local customers for fine optimization. After such processing, the local model parameters finally generated not only retain the broad adaptability of the global model, but also have high-precision prediction capabilities for local markets and customers.

[0086] The model is optimized twice by sending the global model parameters to each training data holder and training and updating the local recursive autoencoder based on these parameters. This process not only ensures that the model retains the global data features, but also makes the model better adapt to the respective data distribution through retraining of local data, thus significantly improving the prediction accuracy and adaptability of the model in practical applications.

[0087] S50, repeatedly performing training, uploading, aggregating and updating operations until a predetermined stop condition is met, stopping the training process, obtaining a target local recursive autoencoder for each training data holder, and obtaining layout data based on the target local recursive autoencoder.

[0088] In this embodiment, through an iterative loop, the training data holder continuously performs the following steps: recursive autoencoder training on local data, uploading local model parameters to the federated server, aggregating model parameters to generate global model parameters, and then updating the local model based on the global model parameters. This cycle is repeated until the preset stop condition is met, such as reaching a specified number of training rounds, the model's loss function converges to a certain threshold, or the model performance is no longer significantly improved. Ultimately, each training data holder obtains an optimized target local recursive autoencoder that can effectively process and generate layout data.

[0089] In one specific implementation, in the financial credit system, each bank optimizes the customer credit scoring model after each round of training and combines it with technical means of time series analysis. Specifically, after each local training, the system not only uploads the generated model parameters to the federal server, but also uploads time series data related to customer credit behavior, such as historical credit score change trends. When aggregating model parameters, the federal server uses these time series data for weighted processing so that the global model parameters more accurately reflect the customer's credit changes. The system repeatedly performs this process until the model's prediction of customer credit risk meets the expected standard, and the target recursive autoencoder model finally generated can provide more dynamic and accurate credit scores, providing stronger support for credit decisions.

[0090] In another specific implementation, in a financial asset management system, an asset management company combines data enhancement technology based on market fluctuations in multiple rounds of training. After each local training is completed, the company not only uploads model parameters, but also uploads market data that has been processed with data enhancement, which simulates asset performance under different market conditions. When aggregating model parameters, the federal server processes these enhanced data so that the generated global model parameters have better market adaptability. The system goes through multiple rounds of training and updating until the model can effectively cope with various market fluctuations. The target recursive autoencoder model finally generated can provide customers with more robust asset allocation recommendations.

[0091] By repeatedly performing training, uploading, aggregation, and updating operations until the predetermined stop condition is met, the model can be continuously optimized using multi-party data under the premise of data privacy protection, and finally a high-precision target recursive autoencoder model is generated. The model has good generalization ability and prediction accuracy in multiple application scenarios, and can effectively generate and process layout data, thereby improving the overall performance of the system and user experience.

[0092] The present invention relates to a layout generation method based on recursive autoencoding, which collects equipment images and equipment status data of underwater target equipment, pre-processes the equipment images, extracts damaged area images of underwater target equipment, extracts damage features from damaged area images and equipment status data, and uses a pre-trained deep learning model to analyze the damage features, generates damage assessment results, and determines the damage level of underwater target equipment. The present invention can realize intelligent detection and rapid assessment of damage to underwater equipment, improves the accuracy and efficiency of damage detection, reduces the subjectivity and errors of manual operation, ensures the timeliness and reliability of equipment status monitoring, and provides a scientific basis for maintenance and insurance claims of underwater equipment.

[0093] In one embodiment, the above S20 includes:

[0094] S201, using a local recursive autoencoder to extract features from the local layout data of each training data holder to obtain a low-dimensional feature vector;

[0095] S202, reconstructing the low-dimensional feature vector through a local recursive autoencoder to generate reconstructed layout data for each local layout data;

[0096] S203, calculating a reconstruction error and a KL divergence based on the local layout data and the corresponding reconstructed layout data, training each local recursive autoencoder, and obtaining local model parameters of each local recursive autoencoder.

[0097] In this embodiment, the recursive autoencoder part compresses the high-dimensional original layout data into a low-dimensional feature vector through layer-by-layer neural network operations. This low-dimensional feature vector retains the core features of the original data while reducing the complexity of the data, allowing the model to perform calculations and analysis more efficiently.

[0098] The task of the decoder is to decode the low-dimensional feature vector generated by the encoder and restore it to a reconstructed layout data with a similar structure to the original input data. The decoder gradually restores the low-dimensional feature vector to high-dimensional data through layer-by-layer deconvolution or linear transformation. The effect of this reconstruction process directly affects the performance of the model because it tests whether the encoder retains enough information during the compression process.

[0099] To optimize the recursive autoencoder model, the system calculates the difference between the original layout data and the reconstructed layout data, i.e., the reconstruction error, which is usually measured by the mean square error (MSE). In addition, the system also calculates the KL divergence, which measures the difference between the probability distribution of the encoder output and the target distribution. These two errors are weighted and summed to form a total loss function, and the model parameters are continuously adjusted through the gradient descent algorithm until the optimal parameter configuration is reached.

[0100] In another specific embodiment, in the insurance claims data analysis system, each insurance company introduced multimodal data fusion technology when training the recursive autoencoder model. In addition to traditional claims data (such as compensation amount, claim reason and event time), the system also combines external data sources such as meteorological data and geographic information. When extracting low-dimensional feature vectors, the encoder integrates the features of multiple data sources, so that the model can take into account more contextual information in claims analysis. These low-dimensional feature vectors are reconstructed into more detailed and accurate claims layout data in the decoder. By calculating the reconstruction error and KL divergence, the system continuously optimizes the model parameters, and the local model parameters finally generated can more effectively identify potential insurance fraud.

[0101] This embodiment can generate high-quality local model parameters by extracting, reconstructing and optimizing local layout data. These parameters show excellent feature extraction and data compression capabilities in a variety of financial application scenarios. The training process of the recursive autoencoder ensures data privacy while greatly improving the adaptability and prediction accuracy of the model, thereby providing strong support for financial institutions in credit scoring, claims analysis and investment management.

[0102] In one embodiment, the above S201 includes:

[0103] S2011, inputting the local layout data of each training data holder into a local recursive autoencoder, and generating a mean and a standard deviation representing the distribution of the local layout data in a latent space;

[0104] S2012, generating a random variable that satisfies a standard normal distribution, and multiplying the random variable by the standard deviation to obtain a disturbance distribution value for adjusting feature representation;

[0105] S2013, adding the disturbance distribution value to the mean value to generate the low-dimensional feature vector, wherein the low-dimensional feature vector contains key feature information of the local layout data.

[0106] In this embodiment, the encoder part of the recursive autoencoder receives local layout data as input and maps these high-dimensional data into the latent space through a multi-layer neural network. The encoder generates the corresponding mean μ and standard deviation σ by learning the intrinsic structure of the layout data. The mean represents the central position of the data distribution in the latent space, while the standard deviation describes the spread or uncertainty of the distribution.

[0107] In order to introduce a certain amount of randomness and enhance the generalization ability of the model, the system generates a random variable τ that follows a standard normal distribution. This random variable is multiplied by the standard deviation σ to generate a perturbation distribution value. This perturbation value is used to adjust the feature representation in the latent space, allowing the model to explore more potential data structures during training.

[0108] Finally, the system adds the generated perturbation distribution value to the mean μ to generate a low-dimensional feature vector z. This low-dimensional feature vector is a compressed data representation that retains the key information in the original layout data while compressing the data from a high-dimensional space to a low-dimensional space. This compressed representation improves computational efficiency while retaining the core feature description of the original data.

[0109] In one specific implementation, the local layout data of each training data holder is processed by a self-attention mechanism before being input into the encoder. Adding a self-attention layer to the input layer of the encoder enables the model to capture the correlation and dependency between different features in the data. The multi-head attention mechanism allows the model to learn information from different representation subspaces, improving the richness and expressiveness of feature extraction. By introducing the attention mechanism, the model can focus on important features more accurately, improve the quality of low-dimensional feature vectors, and thus improve the performance of downstream tasks.

[0110] In another specific embodiment, when generating the perturbation distribution value, reverse KL divergence is used for more stable sampling. Compared with standard normal distribution sampling, reverse KL divergence sampling can reduce the uncertainty of the tail of the Gaussian distribution and provide a more stable and centralized perturbation. Regularization terms are added during the sampling process to prevent overfitting and improve the generalization ability of the model. The improved sampling method makes the generated low-dimensional feature vector more stable and representative, which helps to improve the accuracy of the decoder to reconstruct the data.

[0111] In another specific embodiment, the encoder is divided into multiple residual blocks, each block contains multiple convolutional layers and shortcut connections, allowing information to be transmitted more efficiently in the network. The encoder is divided into multiple residual blocks, each block contains multiple convolutional layers and shortcut connections, allowing information to be transmitted more efficiently in the network. A batch normalization layer is added to the residual block to stabilize the training process and accelerate convergence. The introduction of residual connections enables the encoder to build a deeper network structure, thereby extracting richer and more abstract features and improving the expressiveness of the model.

[0112] This embodiment can effectively compress high-dimensional data while retaining the key features of the original data by extracting features from local layout data and generating low-dimensional feature vectors. This not only improves the computational efficiency of the model, but also enhances the generalization ability of the model, enabling it to better adapt to various complex data scenarios.

[0113] In one embodiment, the above S202 includes:

[0114] S2021, inputting the low-dimensional feature vector into the local recursive autoencoder, using a preset convolution kernel and stride parameter, performing a deconvolution operation on the low-dimensional feature vector to generate a feature map with a higher resolution;

[0115] S2022, applying an activation function to perform a nonlinear transformation on the generated feature map;

[0116] S2023, repeatedly performing deconvolution operations to increase the resolution of the feature map layer by layer, and after each deconvolution operation, continuing to apply the activation function to perform nonlinear transformation until the size of the feature map is completely restored to the same size as the original local layout data;

[0117] S2024, inputting the final feature map generated after multiple deconvolution operations into a linear activation function to convert it into the reconstructed layout data.

[0118] In this embodiment, low-dimensional feature vectors are input into the decoder. The decoder gradually expands these low-dimensional feature vectors into higher-resolution feature maps through deconvolution operations. The deconvolution operation (also known as transposed convolution) expands the data from the low-dimensional space back to a higher dimension by using a preset convolution kernel and stride parameters. This step is to restore the spatial structure of the data so that it gradually approaches the resolution of the original layout data.

[0119] After each deconvolution operation, a nonlinear activation function (such as ReLU or Leaky ReLU) is applied to transform the resulting feature map. These activation functions introduce nonlinear elements, allowing the model to capture more complex features and data patterns, enhancing the model's expressive power. Activation functions are crucial in neural networks because they enable the network to learn and represent nonlinear relationships.

[0120] The resolution of the feature map is increased layer by layer by performing deconvolution and activation operations multiple times. Each operation makes the size of the feature map closer to the size of the original layout data until it is restored to the same dimension as the original data. This method of increasing the resolution layer by layer ensures that the feature map can gradually restore the spatial structure and feature distribution of the original data.

[0121] After the last deconvolution operation, the final feature map enters a linear activation function (such as a linear mapping or a sigmoid function). The linear activation function is used to convert the feature map into reconstructed layout data, ensuring that the output data numerically matches the original input layout data, completing the reconstruction process from low-dimensional vectors to high-dimensional layout data.

[0122] In one specific implementation, a spatial attention mechanism is integrated in the deconvolution operation to enhance the reconstruction of important feature areas. Before each deconvolution operation, a spatial attention module is applied to the feature map to emphasize important areas in the image, such as high-frequency areas or key content. Combining global and local spatial attention ensures that more attention is paid to key areas during the reconstruction process to reduce the impact of background noise. The introduction of the spatial attention mechanism enhances the reconstruction capability of key areas, especially when processing complex layout data, which can effectively improve the reconstruction quality of data.

[0123] In another specific implementation, during the deconvolution process, content-aware convolution is used to dynamically adjust the convolution operation according to the content of the feature map. According to the content of the feature map, the shape and weight of the convolution kernel are automatically adjusted to ensure differentiated processing of different content areas. Combined with deformable convolution, it adapts to geometric deformations in the feature map and improves the ability to reconstruct irregular data structures. Content-aware convolution enables the model to handle complex and diverse content more flexibly when reconstructing layout data, and improves the accuracy and consistency of the reconstructed data.

[0124] This embodiment can effectively reconstruct low-dimensional feature vectors into high-dimensional layout data through multiple deconvolution operations and the application of nonlinear activation functions. It not only ensures the integrity and accuracy of the data, but also captures complex nonlinear relationships, so that the model can adapt to the data reconstruction needs in various application scenarios.

[0125] In one embodiment, in the above S203, the reconstruction error and KL divergence are calculated based on the local layout data and the corresponding reconstructed layout data, and each local recursive autoencoder is trained, including:

[0126] S2031, comparing each data point in the local layout data with the corresponding data point in the reconstructed layout data, and calculating the difference between each pair of data points;

[0127] S2032, squaring the difference between each data point, adding the squared differences of all data points, and dividing the sum by the total number of data points to obtain an average squared error, and using the average squared error as a reconstruction error;

[0128] S2033, extracting distribution parameters in the latent space from the encoder output of the local recursive autoencoder, including a mean and a standard deviation, and calculating the difference between the distribution parameters and a standard normal distribution to obtain a KL divergence;

[0129] S2034, performing a weighted summation of the reconstruction error and the KL divergence to form a total loss function;

[0130] S2035, using a gradient descent algorithm, calculate the gradient of the total loss function relative to the model parameters, and adjust the model parameters of each local recursive autoencoder according to the gradient and the set learning rate.

[0131] In this embodiment, each data point in the local layout data and the reconstructed layout data is mapped one by one. By comparing the difference between the original data point and the reconstructed data point, it is possible to evaluate how much original information the decoder has retained during the reconstruction process. This step is the basis for calculating the reconstruction error, and the accuracy of the reconstruction can be judged by the difference.

[0132] The mean square error (MSE) is calculated by squaring the differences between all data points, then summing these squared differences and dividing by the total number of data points. The mean square error reflects the degree of deviation of the model from the original data during the reconstruction process. The lower the MSE value, the stronger the reconstruction ability of the model.

[0133] The encoder generates distribution parameters in the latent space during feature extraction, mainly including mean μ and standard deviation σ. KL divergence (Kullback-Leibler Divergence) is used to measure the difference between these distribution parameters and the standard normal distribution. By calculating the KL divergence, the encoder of the model can be guided to generate latent variables that are closer to the standard normal distribution, thereby improving the robustness of the model.

[0134] The total loss function is a weighted sum of the reconstruction error and the KL divergence. This function takes into account the difference between the reconstructed data and the original data (measured by MSE) and the difference between the latent space distribution and the standard normal distribution (measured by KL divergence). By weighing the impact of the two, the model is able to generate a reasonable latent variable distribution while maintaining reconstruction accuracy.

[0135] Use the gradient descent algorithm to optimize the model parameters. By calculating the gradient of the total loss function relative to the model parameters, we can determine how to adjust the model parameters to minimize the loss function. The learning rate determines the step size of each adjustment. This process is iteratively optimized and eventually converges the model parameters to the optimal value, thereby improving the model's reconstruction ability and generalization performance.

[0136] In one specific implementation, when calculating the reconstruction error, the system not only uses local layout data and reconstructed layout data, but also combines multimodal data (such as text, image, and audio data) to improve the reconstruction accuracy by fusing data features of different modalities. Before calculating the reconstruction error, the features from different data sources are first fused through the multimodal fusion layer. The fused features are then compared with the reconstructed data. Multimodal data fusion improves the reconstruction accuracy and enables the model to better adapt to complex data scenarios, such as financial document analysis that processes image and text information at the same time.

[0137] In another specific implementation, when calculating the reconstruction error, the system uses locality sensitive hashing (LSH) technology to accelerate the similarity comparison between high-dimensional data points and reduce the computational overhead. The data points are mapped to a low-dimensional hash space through LSH, and the reconstruction error calculation is performed in the space to reduce the computational complexity. After calculating the reconstruction error in the LSH space, an error adjustment mechanism is applied to ensure that the error of the hash mapping does not affect the final reconstruction accuracy. Locality sensitive hashing technology significantly reduces the overhead of calculating the reconstruction error of high-dimensional data while maintaining a high computational accuracy, which is suitable for processing ultra-large-scale financial data sets.

[0138] In other specific embodiments, when calculating KL divergence, variational Bayesian inference is used to more accurately estimate the distribution of the latent space and reduce the reliance on KL divergence. A variational Bayesian inference is used to more accurately estimate the distribution of the latent space output by the encoder, alleviating the limitations of the standard normal distribution. By optimizing the variational lower bound (Evidence Lower Bound, ELBO), the fitting ability for complex distributions is improved and the impact of KL divergence is reduced. Variational Bayesian inference enables the model to better handle non-Gaussian distributed latent spaces, optimizes the calculation process of KL divergence, and is suitable for financial data with complex distribution characteristics.

[0139] This embodiment can optimize the distribution of potential variables while maintaining the model reconstruction accuracy by calculating the reconstruction error and KL divergence. The gradient descent algorithm is used to optimize the model parameters, so that the recursive autoencoder can provide high-quality feature extraction and data reconstruction capabilities in a variety of financial application scenarios, thereby improving the prediction accuracy and stability of the model.

[0140] In one embodiment, in the above S30, the local model parameters are aggregated by the federal server to generate global model parameters, including:

[0141] S301, calculating the weight of each local recursive autoencoder according to the ratio of the amount of local layout data of each training data holder to the total amount of local layout data of all training data holders;

[0142] S302, based on the weight of each local recursive autoencoder, multiply the local model parameters of each local recursive autoencoder by the corresponding weight through the federal server, and sum all the weighted model parameters to finally generate the global model parameters.

[0143] In this embodiment, the federated server first counts the amount of local layout data of each training data holder, compares it with the total amount of data of all participants, and calculates the proportion of each training data holder in the entire data set. This proportion is used as the weight of the local recursive autoencoder, reflecting the contribution of each participant in the global model parameter aggregation process.

[0144] The federated server multiplies the local model parameters of each training data holder by their corresponding weights to obtain the weighted model parameters. Then, the federated server sums up all the weighted model parameters to form the final global model parameters. This global model parameter integrates the local model information of all participants, and the influence of each participant on the global model is proportional to the amount of data.

[0145] In one specific implementation, when aggregating model parameters, a method based on weighted average is used, and dynamic adjustment of weights by historical model parameters is further introduced. When calculating the weight of each local recursive autoencoder, not only the current amount of data is considered, but also factors of historical model performance and convergence speed are introduced. By considering historical data and dynamically adjusting weights, more attention is paid to the long-term performance of the model during the aggregation process. During multiple rounds of training, an attenuation mechanism is applied to the weights of each participant to reduce dependence on a single data source and improve the robustness of the model. It ensures that the model can find a balance between different amounts of data and historical performance, thereby improving the robustness and long-term performance of the global model during aggregation.

[0146] In another specific implementation, when calculating the weights of local model parameters, a differential privacy mechanism is used to protect the data privacy of the participants. When calculating the weights, the federated server adds noise to the local data volume of each participant to prevent the actual data distribution and scale of the participant from being inferred through the data volume. Through privacy budget control, it is ensured that the sensitive information of the participants is not leaked during the calculation of the weights, and the privacy protection and model performance are balanced. The differential privacy mechanism improves the security of the entire federated learning system, ensures that data privacy is not leaked, and maintains the effectiveness of model parameter aggregation.

[0147] This embodiment aggregates local model parameters on the federated server to generate global model parameters, effectively integrating the local model information of each training data holder and improving the overall performance of the global model. By calculating the weights and weighted summing the model parameters of each party, it ensures that participants with larger data volumes and more contributions have greater influence in the global model, while also ensuring that the global model can make full use of multi-source data to achieve higher accuracy and robustness. In addition, while maintaining the data privacy of each participant, the fusion process of the model parameters is optimized, significantly improving the efficiency and security of the federated learning system.

[0148] In one embodiment, in the above S40, training the local recursive autoencoder based on the global model parameters and updating the local model parameters of each local recursive autoencoder includes:

[0149] S401, using the global model parameters as initial model parameters of the local recursive autoencoder, and combining with the local layout data, calculating the loss function of the initial model parameters on the local layout data;

[0150] S402, calculating the gradient of the loss function equivalent to the initial model parameters through a gradient descent algorithm;

[0151] S403, setting a learning rate for the local recursive autoencoder, multiplying the gradient by the learning rate, and generating an update value for adjusting the initial model parameters;

[0152] S404: Adjust the initial model parameters based on the update value, and update the local model parameters of each local recursive autoencoder.

[0153] In this embodiment, the global model parameters generated by the federated server are sent to each training data holder and used as the initial model parameters of the local recursive autoencoder. These initial model parameters are combined with the local layout data of each holder to evaluate the performance of the current model parameters by calculating the loss function. The loss function usually measures the error between the model prediction and the actual data. By minimizing the loss function, the model can be gradually optimized.

[0154] After calculating the loss function, the system uses the gradient descent algorithm to calculate the gradient of the loss function relative to the model parameters. The gradient indicates the direction and speed of change of the loss function, and guides the adjustment direction of the model parameters. By calculating this gradient, the model can identify how to adjust the parameters to reduce the error.

[0155] The learning rate determines the step size of each model parameter update. Multiply the gradient by the learning rate to generate an update value used to adjust the model parameters. A learning rate that is too high may cause unstable training, while a learning rate that is too low may make the training process too slow.

[0156] The calculated update values ​​are used to adjust the initial model parameters. The adjusted model parameters are updated in the local recursive autoencoder, making the model better adapted to the local layout data. Through successive iterations, the local model parameters are continuously optimized, improving the performance of the model on local data.

[0157] In one specific implementation, during the model parameter update process, the system dynamically adjusts the learning rate according to the current training stage and error changes to improve training efficiency and model stability. A dynamic learning rate scheduler is introduced to automatically adjust the learning rate according to the convergence speed of the loss function and the training stage to prevent overfitting or slow convergence. During the training process, adaptive learning rates are applied to different parameters or levels to further optimize the convergence and generalization ability of the model. The dynamic learning rate adjustment technology improves the efficiency of the training process, avoids overfitting and underfitting, and enables the model to adapt to local data more quickly and accurately.

[0158] In another specific embodiment, during the gradient descent process, the system introduces gradient clipping technology to avoid training instability caused by gradient explosion. After the gradient is calculated, the gradient value is clipped so that it does not exceed the preset threshold, thereby maintaining the stability of the training. According to the different stages of model training, the gradient clipping strategy is dynamically adjusted to ensure that the clipping process can prevent gradient explosion without affecting the training effect of the model. Gradient clipping technology improves the stability of the training process, especially when processing deep neural networks or high-dimensional data, effectively avoiding the problem of training out of control caused by gradient explosion. .

[0159] This embodiment significantly improves the adaptability and performance of the model in different data environments by using global model parameters to guide the training and parameter update of local recursive autoencoders. By calculating the loss function, gradient, and setting the learning rate, the system can effectively optimize the local model parameters to ensure that the improvement of the global model can be fully reflected in the local data of each participant.

[0160] The present invention also provides a layout generation device based on recursive self-encoding, referring to Figure 2 , Figure 2 The functional module diagram of a preferred embodiment of the layout generation device based on recursive self-encoding of the present invention is shown in FIG. The layout generation device based on recursive self-encoding comprises:

[0161] A local data acquisition module, which obtains local layout data from multiple training data holders respectively, wherein the local layout data includes position information, size information, and type information of a bounding box;

[0162] A local model training module trains the local recursive autoencoder of each training data holder based on each local layout data to obtain local model parameters of each local recursive autoencoder;

[0163] A model aggregation module uploads all local model parameters to a federal server, aggregates the local model parameters through the federal server, and generates global model parameters;

[0164] A local model updating module, which sends the global model parameters to each training data holder through the federal server, trains a local recursive autoencoder based on the global model parameters, and updates the local model parameters of each local recursive autoencoder;

[0165] The iterative training module repeatedly performs training, uploading, aggregation and updating operations until a predetermined stopping condition is met, stops the training process, obtains a target local recursive autoencoder for each training data holder, and obtains layout data based on the target local recursive autoencoder.

[0166] The specific implementation of the layout generation device based on recursive self-encoding of the present invention is basically the same as the above-mentioned embodiments of the layout generation method based on recursive self-encoding, and will not be repeated here.

[0167] The present invention also provides a layout generation device based on recursive self-encoding, such as Figure 3 As shown, the layout generation device based on recursive self-encoding may include: a processor 1001, such as a CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or it may be a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0168] Those skilled in the art will understand that Figure 3 The hardware structure of the layout generation device based on recursive self-encoding shown in the figure does not constitute a limitation of the layout generation device based on recursive self-encoding, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0169] like Figure 3As shown, the memory 1005 as a storage medium may include an operating system, a network communication module, a user interface module, and a layout generation program based on recursive self-encoding. The operating system is a program for managing and controlling the layout generation device and software resources based on recursive self-encoding, and supports the operation of the network communication module, the user interface module, the layout generation program based on recursive self-encoding, and other programs or software; the network communication module is used to manage and control the network interface 1004; the user interface module is used to manage and control the user interface 1003.

[0170] exist Figure 3 In the hardware structure of the layout generation device based on recursive self-encoding shown, the network interface 1004 is mainly used to connect to the background server and communicate data with the background server; the user interface 1003 is mainly used to connect to the client and communicate data with the client; the processor 1001 can call the layout generation program based on recursive self-encoding stored in the memory 1005, and perform the same operation as the layout generation method based on recursive self-encoding.

[0171] The specific implementation of the layout generation device based on recursive self-encoding of the present invention is basically the same as the above-mentioned embodiments of the layout generation method based on recursive self-encoding, and will not be repeated here.

[0172] In addition, an embodiment of the present invention further proposes a computer storage medium, on which a layout generation program based on recursive self-encoding is stored. When the layout generation program based on recursive self-encoding is executed by a processor, the steps of the layout generation method based on recursive self-encoding as described above are implemented.

[0173] The specific implementation of the computer storage medium of the present invention is basically the same as the above-mentioned embodiments of the layout generation method based on recursive self-encoding, and will not be repeated here.

Claims

1. A layout generation method based on recursive autoencoding, characterized in that: The following steps are involved: Acquire local layout data from multiple training data holders respectively, where the local layout data includes location information, size information, and type information of a bounding box; Based on each local layout data, a local recursive autoencoder of each training data holder is trained to obtain a local model parameter of each local recursive autoencoder; Upload all local model parameters to the federal server, aggregate the local model parameters through the federal server, and generate global model parameters; Sending the global model parameters to each training data holder through the federal server, training a local recursive autoencoder based on the global model parameters, and updating the local model parameters of each local recursive autoencoder; The training, uploading, aggregating and updating operations are repeatedly performed until a predetermined stopping condition is met, the training process is stopped, a target local recursive autoencoder of each training data holder is obtained, and layout data is obtained based on the target local recursive autoencoder.

2. The layout generation method based on recursive autoencoding according to claim 1, characterized in that: The local recursive autoencoder of each training data holder is trained based on each local layout data to obtain local model parameters of each local recursive autoencoder, including: Use a local recursive autoencoder to extract features from the local layout data of each training data holder to obtain a low-dimensional feature vector; Reconstructing the low-dimensional feature vector through a local recursive autoencoder to generate reconstructed layout data for each local layout data; The reconstruction error and KL divergence are calculated based on the local layout data and the corresponding reconstructed layout data, and each local recursive autoencoder is trained to obtain the local model parameters of each local recursive autoencoder.

3. The layout generation method based on recursive autoencoding according to claim 2, characterized in that: Use the local recursive autoencoder to extract features from the local layout data of each training data holder to obtain a low-dimensional feature vector, including: Input the local layout data of each training data holder into the local recursive autoencoder and generate the mean and standard deviation representing the distribution of the local layout data in the latent space; Generate a random variable that satisfies a standard normal distribution, and multiply the random variable by the standard deviation to obtain a perturbation distribution value for adjusting feature representation; The disturbance distribution value is added to the mean value to generate the low-dimensional feature vector, which contains key feature information of the local layout data.

4. The layout generation method based on recursive autoencoding as described in claim 2, characterized in that: The low-dimensional feature vector is reconstructed through the local recursive autoencoder to generate reconstructed layout data for each local layout data, including: Inputting the low-dimensional feature vector into the local recursive autoencoder, and performing a deconvolution operation on the low-dimensional feature vector using a preset convolution kernel and stride parameter to generate a feature map with a higher resolution; Apply activation function to perform nonlinear transformation on the generated feature map; Repeat the deconvolution operation to increase the resolution of the feature map layer by layer. After each deconvolution operation, continue to apply the activation function for nonlinear transformation until the size of the feature map is completely restored to the same size as the original local layout data; The final feature map generated after multiple deconvolution operations is input into a linear activation function and converted into the reconstructed layout data.

5. The layout generation method based on recursive autoencoding according to claim 2, characterized in that: The reconstruction error and KL divergence are calculated based on the local layout data and the corresponding reconstructed layout data, and each local recursive autoencoder is trained, including: Compare each data point in the local layout data with the corresponding data point in the reconstructed layout data, and calculate the difference between each pair of data points; The difference between each data point is squared, and the squared differences of all data points are added together and divided by the total number of data points to obtain an average square error, which is used as the reconstruction error; Extract the distribution parameters in the latent space, including the mean and standard deviation, from the encoder output of the local recursive autoencoder, and calculate the difference between the distribution parameters and the standard normal distribution to obtain the KL divergence; Performing a weighted summation of the reconstruction error and the KL divergence to form a total loss function; Using the gradient descent algorithm, the gradient of the total loss function with respect to the model parameters is calculated, and the model parameters of each local recursive autoencoder are adjusted by the gradient and the set learning rate.

6. The layout generation method based on recursive autoencoding according to claim 1, characterized in that: The local model parameters are aggregated by the federal server to generate global model parameters, including: Calculate the weight of each local recursive autoencoder according to the ratio of the amount of local layout data of each training data holder to the total amount of local layout data of all training data holders; Based on the weight of each local recursive autoencoder, the local model parameters of each local recursive autoencoder are multiplied by the corresponding weight through the federal server, and all weighted model parameters are summed to finally generate the global model parameters.

7. The layout generation method based on recursive autoencoding according to claim 1, characterized in that: Training the local recursive autoencoder based on the global model parameters and updating the local model parameters of each local recursive autoencoder includes: Using the global model parameters as initial model parameters of the local recursive autoencoder, and combining with the local layout data, calculating the loss function of the initial model parameters on the local layout data; Calculate the loss function equivalent to the gradient of the initial model parameters through the gradient descent algorithm; Setting a learning rate for the local recursive autoencoder, multiplying the gradient by the learning rate to generate an update value for adjusting the initial model parameters; The initial model parameters are adjusted based on the updated values, and the local model parameters of each local recursive autoencoder are updated.

8. A layout generation device based on recursive autoencoding, characterized in that: The layout generation device based on recursive autoencoding comprises: A local data acquisition module, which obtains local layout data from multiple training data holders respectively, wherein the local layout data includes position information, size information, and type information of a bounding box; A local model training module trains the local recursive autoencoder of each training data holder based on each local layout data to obtain local model parameters of each local recursive autoencoder; A model aggregation module uploads all local model parameters to a federal server, aggregates the local model parameters through the federal server, and generates global model parameters; A local model updating module, which sends the global model parameters to each training data holder through the federal server, trains a local recursive autoencoder based on the global model parameters, and updates the local model parameters of each local recursive autoencoder; The iterative training module repeatedly performs training, uploading, aggregation and updating operations until a predetermined stopping condition is met, stops the training process, obtains a target local recursive autoencoder for each training data holder, and obtains layout data based on the target local recursive autoencoder.

9. A layout generation device based on recursive autoencoding, characterized in that: The recursive autoencoding-based layout generation device includes a memory, a processor, and a recursive autoencoding-based layout generation program stored in the memory and executable on the processor. When the recursive autoencoding-based layout generation program is executed by the processor, the steps of the recursive autoencoding-based layout generation method as described in any one of claims 1 to 7 are implemented.

10. A computer storage medium, characterized in that: The storage medium stores a layout generation program based on recursive self-encoding, and when the layout generation program based on recursive self-encoding is executed by a processor, the steps of the layout generation method based on recursive self-encoding as described in any one of claims 1 to 7 are implemented.

Images