Data generation method and device based on reversible transformation parameters, equipment and medium

Abstract

The invention relates to the technical field of artificial intelligence, can be applied to business scenes such as financial science and technology and medical health, and discloses a data generation method, device and equipment based on reversible transformation parameters and a medium. And dynamically generating reversible transformation parameters of the flow model based on the aggregation characteristics, sampling to generate an initial hidden variable and taking the initial hidden variable as current input data, executing reversible linear transformation, segmentation, transformation and coupling operations, and repeating the processes until output data is generated. According to the method, the reversible transformation parameters of the flow model are dynamically generated based on the condition information, the super-network structure is combined with the multi-layer reversible transformation process, flexible control over different condition information is achieved, the diversity and adaptation capacity of condition data generation are improved, limitation caused by dependence on fixed structure design is avoided, and the method is suitable for large-scale popularization and application. And the method has higher generation flexibility and generalization performance.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a data generation method, apparatus, device, and storage medium based on reversible transformation parameters. Background Technology

[0002] Existing conditional generation methods are widely used within the framework of normalizing flow models, demonstrating good performance, especially in complex data generation tasks such as image, speech, financial, and healthcare data. Through multi-layered reversible transformation structures, flow models can effectively characterize data distributions and support rigorous likelihood estimation. However, in specific conditional data generation tasks, existing techniques generally suffer from weak expression of conditional information, complex design of conditional control structures, and a lack of dynamic adjustment mechanisms for reversible transformation parameters, severely limiting the applicability and generation effectiveness of the models.

[0003] Specifically, existing methods typically introduce conditional information through explicit structural modifications. Common techniques include embedding independent conditional input paths into the network structure or designing separate parameter modules for different condition types. While this approach achieves conditional control to some extent, it significantly increases the complexity of model design, leading to a rigid overall architecture, inefficient parameters, and difficulty in adapting to varying conditional input requirements. Furthermore, conditional information usually acts directly on the input layer or parts of the shallow structure, lacking high-order representations and dynamic relationships for deep features. This makes it difficult to capture the deep relationships between complex conditions and data distribution, impacting the diversity and accuracy of the generated data.

[0004] Furthermore, in the reversible transformation process of multi-layered stacking in flow models, parameters are typically fixed, lacking the ability to dynamically adjust reversible transformation parameters based on specific conditions. Especially in tasks involving high-dimensional data, complex conditions, or cross-domain data generation, fixed parameter design limits the model's ability to characterize complex data distributions and also affects the efficient control of the generated results by conditional information. Existing technologies still have significant shortcomings in multi-level conditional control and dynamic parameter generation, making it difficult to simultaneously achieve flexible conditional control, scalability of the model structure, and efficient reversibility in the generation process.

[0005] In the fintech business field, the synthesis and enhancement of structured or semi-structured business data relies on efficient condition control and dynamic generation capabilities. However, existing conditional flow models cannot flexibly adjust reversible transformation parameters for different business rules, risk factors, or customer profiles, which affects the authenticity of simulated data and the accuracy of risk modeling.

[0006] In the field of healthcare, when faced with the need to generate multimodal health data, such as specific data simulation based on individual health indicators, disease labels, or clinical conditions, existing technologies lack high-order condition expression and dynamic parameter generation mechanisms. This makes it difficult to ensure an accurate correspondence between the generated data and complex health conditions, thus restricting the application effectiveness of scenarios such as personalized medical simulation, disease prediction, and assisted diagnosis. Summary of the Invention

[0007] The main objective of this invention is to provide a data generation method, apparatus, device, and storage medium based on reversible transformation parameters. This invention aims to solve the technical problem that the existing technology lacks a mechanism for dynamically generating reversible transformation parameters by combining conditional information, which leads to the inability of the flow model to flexibly adapt to different conditions during multi-layer transformation, thus limiting the accuracy of conditional control and the generation effect.

[0008] To achieve the above objectives, the present invention provides a data generation method based on reversible transformation parameters, comprising:

[0009] Obtain condition information and encode the condition information to obtain a condition vector;

[0010] The conditional vector is input into the multi-layer feature encoder in the hypernetwork generation module to extract the high-order features of the conditional vector and obtain the aggregated features.

[0011] The aggregated features are input into the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model;

[0012] Initial latent variables are generated by sampling from a preset distribution, and these initial latent variables are used as the current input data;

[0013] Perform an invertible linear transformation on the current input data to generate the transformed input data;

[0014] The transformed input data is divided into a first data part and a second data part;

[0015] Using the scale parameter and translation parameter in the reversible transformation parameters, a scale transformation and a translation transformation are performed on the second data part to generate the transformed second data part;

[0016] The first data portion is coupled with the transformed second data portion to generate new current input data;

[0017] Repeatedly perform reversible linear transformation, segmentation, transformation and coupling operations on the new current input data until output data is generated.

[0018] Furthermore, to achieve the above objectives, the present invention provides a data generation apparatus based on reversible transformation parameters, comprising:

[0019] A condition encoding module is used to acquire condition information and encode the condition information to obtain a condition vector;

[0020] The feature extraction module is used to input the conditional vector into the multi-layer feature encoder in the hypernetwork generation module, extract the high-order features of the conditional vector, and obtain aggregated features;

[0021] The parameter generation module is used to input the aggregated features into the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model;

[0022] The latent variable initialization module is used to sample and generate initial latent variables from a preset distribution, and use the initial latent variables as the current input data;

[0023] The reversible transformation module is used to perform a reversible linear transformation on the current input data to generate the transformed input data;

[0024] The data segmentation module is used to segment the transformed input data into a first data part and a second data part.

[0025] The data transformation module is used to perform scale transformation and translation transformation on the second data part using the scale parameter and translation parameter in the reversible transformation parameters to generate the transformed second data part;

[0026] A data coupling module is used to couple the first data portion with the transformed second data portion to generate new current input data;

[0027] The multi-level iterative module is used to repeatedly perform reversible linear transformation, segmentation, transformation and coupling operations on the new current input data until output data is generated.

[0028] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a data generation program based on reversible transformation parameters stored in the memory and executable on the processor, wherein when the data generation program based on reversible transformation parameters is executed by the processor, it implements the steps of the data generation method based on reversible transformation parameters as described above.

[0029] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a data generation program based on reversible transformation parameters, wherein the data generation program based on reversible transformation parameters, when executed by a processor, implements the steps of the data generation method based on reversible transformation parameters as described above.

[0030] Beneficial Effects: This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses a data generation method, apparatus, device, and medium based on reversible transformation parameters, comprising: acquiring conditional information and encoding it to obtain a conditional vector; inputting the conditional vector into a multi-layer feature encoder within a hypernetwork generation module; extracting high-order features from the conditional vector to obtain aggregated features; dynamically generating reversible transformation parameters for a flow model based on the aggregated features; sampling and generating initial latent variables from a preset distribution as current input data; performing a reversible linear transformation on the current input data to generate transformed input data; segmenting the transformed input data into a first data part and a second data part; performing a scale transformation and a translation transformation on the second data part using the scale and translation parameters in the reversible transformation parameters to generate the transformed second data part; coupling the first data part and the transformed second data part to generate new current input data; and repeatedly performing the reversible linear transformation, segmentation, transformation, and coupling operations until output data is generated. This invention combines a supernetwork structure with a multi-layer reversible transformation process by dynamically generating reversible transformation parameters of a flow model based on conditional information. This enables flexible control for different conditional information, improves the diversity and adaptability of conditional data generation, avoids the limitations of relying on fixed structure design in traditional methods, and has higher generation flexibility and generalization performance. Attached Figure Description

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0032] Figure 1 This is a schematic diagram of an application environment for a data generation method based on reversible transformation parameters according to an embodiment of the present invention;

[0033] Figure 2 This is a flowchart illustrating an embodiment of the data generation method based on reversible transformation parameters according to the present invention.

[0034] Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the data generation device based on reversible transformation parameters of the present invention;

[0035] Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0036] Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0037] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0038] The data generation method based on reversible transformation parameters provided in this invention can be applied to, for example... Figure 1 In this application environment, the user terminal communicates with the server via a network. The server can obtain conditional information from the user terminal and encode it into a conditional vector. This conditional vector is then input into a multi-layer feature encoder within the hypernetwork generation module to extract high-order features from the conditional vector, obtaining aggregated features. Based on these aggregated features, reversible transformation parameters for the flow model are dynamically generated. Initial latent variables are sampled from a preset distribution and used as the current input data. A reversible linear transformation is performed on the current input data to generate transformed input data. The transformed input data is then segmented into a first data part and a second data part. The scale and translation parameters in the reversible transformation parameters are used to perform a scale transformation and a translation transformation on the second data part to generate the transformed second data part. The first data part and the transformed second data part are coupled to generate new current input data. This process of reversible linear transformation, segmentation, transformation, and coupling is repeated until output data is generated. This invention combines a hypernetwork structure with a multi-layer reversible transformation process by dynamically generating reversible transformation parameters of a flow model based on conditional information. This enables flexible control over different conditional information, enhancing the diversity and adaptability of conditional data generation. It avoids the limitations of traditional methods that rely on fixed structure designs, resulting in higher generation flexibility and generalization performance. The user end can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server end can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0039] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the data generation method based on reversible transformation parameters provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0040] like Figure 2 As shown, the data generation method based on reversible transformation parameters proposed in this invention includes the following steps:

[0041] S10, Obtain condition information and encode the condition information to obtain a condition vector;

[0042] In this embodiment, the process of acquiring conditional information involves obtaining data content with clear business meaning or environmental information through data interfaces, data storage structures, or external communication modules. Conditional information may include category labels, environmental variables, text descriptions, semantic identifiers, task instructions, scene configurations, historical state information, or other distinctive or guiding input information. Category labels may originate from classification fields in a structured database; environmental variables may originate from a real-time sensing system; text descriptions may originate from user interaction input or upstream system text output; semantic identifiers may be generated through rule extraction or a predefined dictionary; task instructions may originate from the process engine control module; scene configurations may be loaded based on the system initialization configuration file; and historical state information may originate from a log system, cache system, or state database. The source of conditional information is not limited to a single structure or source channel, supporting cross-system, cross-module, and cross-scene data integration methods.

[0043] The process of encoding conditional information involves transforming raw conditional information of different forms, dimensions, and data types into a unified numerical representation structure, facilitating subsequent computation and feature modeling. Encoding operations can be implemented in various ways depending on the data type, business meaning, or input requirements of the downstream model. For category labels, encoding operations can include one-hot encoding, label mapping, sparse matrix expansion, or hash transformation. For text descriptions, encoding operations can be implemented using word vector models, pre-trained language models, character-level embeddings, sub-word segmentation, or sequence encoding networks. For other types of data such as environmental variables, semantic identifiers, task instructions, scene configurations, and historical state information, encoding operations can be implemented based on rule mapping, numerical normalization, bucketing mapping, discretization transformation, continuous compression, or low-dimensional representation expansion. The ultimate goal of encoding operations is to transform conditional information into a numerical vector with fixed dimensions, continuous expression, differentiable structure, and that meets the requirements of downstream computation.

[0044] The process of obtaining condition vectors refers to the formation of a multidimensional array or tensor structure with a unified structure, fixed dimensions, and standard numerical expression format after the encoding operation. The dimensions, data type, and numerical range of the condition vectors can be configured according to actual needs, typically using floating-point numbers. The number of dimensions can be dynamically adjusted based on the richness of the conditional information, information capacity requirements, or model structure design. Condition vectors can adopt the structure of one-dimensional arrays, two-dimensional matrices, or high-dimensional tensors to easily adapt to the input requirements of different computation modules. As one of the inputs in subsequent computation processes, the condition vectors must possess good numerical stability, information integrity, and accuracy of expression.

[0045] In implementing different business scenarios, conditional information can be obtained by calling external interfaces to receive structured classification label data, by reading natural language text information input by users, or by obtaining environmental variable information in real time through the system status monitoring module. Conditional information can be data from a single source, or it can be formed by fusing multiple data sources to create a composite conditional information structure containing various types of information.

[0046] In the process of encoding conditional information, for category labels, a one-hot encoding scheme can be used to map category information into a sparse vector of fixed dimensions; for text descriptions, a pre-trained BERT model or a similar language model can be used to map text information into a context-dependent semantic vector of fixed dimensions; for environmental variables, normalization mapping can be used to convert data of different dimensions into a unified numerical range expression; for complex type conditional information, a multi-branch encoding network can be used to encode different types of information independently, and finally a unified conditional vector output can be formed through vector concatenation, weighted fusion, or interactive attention mechanisms.

[0047] In obtaining condition vectors, the dimensional design of the condition vectors can be dynamically adjusted according to the model size and business complexity, depending on the needs of different industries. For example, in text generation tasks, condition vectors can be designed as 512-dimensional floating-point tensors, while in image generation tasks, they can be designed as 1024-dimensional or higher-dimensional floating-point tensors to meet the needs of more complex condition information representation.

[0048] Example Explanation: In the healthcare business field, the process of obtaining conditional information can be achieved by accessing an electronic medical record system to acquire the patient's structured diagnostic tags, past medical history text descriptions, or real-time physiological monitoring data. Encoding operations on the conditional information can include mapping diagnostic tags to one-hot encoded vectors, converting past medical history texts into semantic vectors using a language model, and normalizing and mapping physiological monitoring data into numerical vectors, ultimately obtaining a multi-dimensional conditional vector expressing the patient's current health status, diagnostic and treatment information, and real-time monitoring data.

[0049] In the fintech business, acquiring conditional information can be achieved by connecting to user authentication systems, transaction behavior log systems, or risk assessment systems to obtain user identity category tags, transaction behavior text descriptions, or risk level indicator data. Encoding operations for this conditional information can include mapping identity tags to one-hot encoded vectors, converting transaction behavior text into behavioral feature vectors using a language model, and standardizing risk indicator data into numerical representations. The final result is a conditional vector containing user identity information, transaction behavior characteristics, and risk level indicators, facilitating efficient processing by subsequent generation and decision-making systems.

[0050] This embodiment, through the aforementioned operational process, effectively achieves unified acquisition and expression of conditional information of multiple types, sources, and structures, avoiding the information fragmentation and inconsistent expression problems caused by the diverse types and complex structures of conditional information in traditional systems. Through an efficient, flexible, and scalable encoding mechanism, it ensures the fusion and expression of different types of conditional information in a unified numerical space, enhancing the downstream system's ability to express and model conditional information. The structured output of conditional vectors ensures the integrity of information expression, the stability of numerical expression, and the efficiency of the computation process, further improving the overall system performance and reliability.

[0051] S20, the conditional vector is input into the multi-layer feature encoder in the hypernetwork generation module to extract the high-order features of the conditional vector and obtain the aggregated features;

[0052] In this embodiment, the operation of inputting conditional vectors into the multi-layer feature encoder within the hypernetwork generation module refers to transferring the encoded conditional vectors, which possess a unified numerical structure and fixed-dimensional expression, to the predefined hypernetwork generation module via a data flow transmission mechanism. The hypernetwork generation module is a computational structure with dynamic parameter generation capabilities, incorporating a multi-layer feature encoder that provides computational functions for progressively abstracting and enhancing the expressive power of input features. The multi-layer feature encoder typically consists of multiple independent but interconnected encoding layers. Each encoding layer progressively improves the expressive dimension and information abstraction level of the input features through operations such as local parameter transformation, structural information enhancement, and global information capture. The input operation includes not only simple data transmission but also verification and standardization of the input data format, dimension, and numerical range to ensure that the input data meets the computational requirements of the encoder structure.

[0053] The process of extracting higher-order features from conditional vectors involves a multi-layer feature encoder that, through multiple nonlinear transformations, parameter mappings, information interactions, and structural enhancements, gradually captures more complex, abstract, and information-integrating higher-order representations from the input conditional vectors. Specifically, each encoding layer can include a multi-head self-attention structure to capture long-distance dependencies and global information associations in the input features; a feedforward neural network structure to enhance nonlinear expressive power; and normalization and residual connection structures to improve computational stability and information retention. The essence of higher-order features is a multi-dimensional feature representation that, compared to the input conditional vector, possesses stronger information expressive power, higher information compression density, and richer interactive information.

[0054] The process of obtaining aggregated features refers to the process of combining multiple sets of intermediate features output by a multi-layer feature encoder through weighted fusion, feature concatenation, attention weighting, or other information integration mechanisms to ultimately form a feature result with a single expressive structure. Aggregated features typically possess attributes such as fixed structure, uniform numerical format, complete expressive information, and sufficient information correlation, facilitating further dynamic parameter generation or conditional information control by subsequent modules. Aggregated features not only retain the basic expressive content of the input conditional information but also integrate rich information from different encoding levels and different expression granularities through multi-layer abstraction and information fusion, forming a feature with high expressive power that is superior to the basic input in overall expressive performance.

[0055] In different business implementation environments, the operation of inputting conditional vectors into the multi-layer feature encoder within the hypernetwork generation module can be implemented based on data transmission bus, memory pointer passing, parameter binding calls, or distributed data transmission mechanisms. For small-scale data processing, direct function calls and memory pointer passing can be used to complete the data input operation; for large-scale data or distributed deployment scenarios, remote procedure calls, parameter synchronization mechanisms, or cross-node data buffering mechanisms can be used to complete the data input.

[0056] In the structural design of multi-layer feature encoders, one can base it on the standard Transformer encoder stacking structure, where each encoding layer includes a multi-head self-attention module, a feedforward neural network module, a residual connection structure, and a normalization structure. Alternatively, one can base it on lightweight encoder structures, such as improved Transformers, hierarchical self-attention structures, or local receptive field enhancement structures, to improve the balance between computational efficiency and expressive power.

[0057] In the process of high-order feature extraction, multi-layer feature encoders can be configured with different numbers of layers, parameter scales, and computational structures to meet different task requirements. With more layers, richer global information and long-distance dependencies can be captured, making them suitable for complex conditional information representation needs; with fewer layers, computational efficiency can be improved, making them suitable for scenarios with high real-time requirements.

[0058] In the process of obtaining aggregated features, various information fusion methods, such as weighted averaging, attention weighting, feature concatenation, convolutional fusion, or dynamic gating, can be used to integrate multi-layer output features. The information fusion mechanism can be flexibly adjusted according to the input information structure, task expression requirements, and downstream module input requirements to ensure the structural integrity and expressive effectiveness of the aggregated features.

[0059] Example Explanation: In the healthcare business domain, the process of inputting conditional vectors into the multi-layer feature encoder within the hypernetwork generation module can be achieved through a system interface. This interface integrates the patient's diagnostic label vector, past medical history vector, and physiological monitoring data into a conditional vector, which is then input into the encoder structure. The multi-layer feature encoder, through multi-layer information abstraction and interaction, extracts high-order representations of the patient's overall health information, forming aggregated features for subsequent medical decision support, diagnosis generation, or health status prediction.

[0060] In the fintech business, conditional vectors can be integrated with user identity tags, transaction behavior characteristics, and risk assessment indicators to form inputs, which are then fed into a multi-layer feature encoder structure. The encoder, through information interaction and expression enhancement, extracts deep-seated correlation features of user financial behavior, obtaining aggregated features that express user behavior patterns and risk characteristics. This facilitates subsequent system implementation of personalized financial product generation, risk prediction, and dynamic adjustment of compliance control strategies.

[0061] This embodiment, through the aforementioned operations, effectively enhances the extraction and expression capabilities of input conditional information at multiple levels, angles, and granularities, overcoming the problems of limited expression dimensions and insufficient information correlation in basic input information. The structural design of the multi-layer feature encoder and the dynamic information fusion mechanism ensure a balance between the richness of high-order feature expression, information completeness, and computational efficiency, improving the overall system's ability to model and express complex conditional information accurately. The resulting aggregated features possess excellent information expression density and information integration capabilities, significantly enhancing the dynamic parameter generation effect and conditional information control capability of the subsequent system.

[0062] S30, the aggregated features are input to the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model;

[0063] In this embodiment, the operation of inputting aggregated features into the gated parameter projection layer refers to the process of transferring aggregated features, which have undergone information extraction and multi-level expression integration, to the gated parameter projection layer structure, which possesses parameter mapping and dynamic generation capabilities, through data transmission and computation. Aggregated features typically possess uniform dimensions, standard numerical ranges, and multi-dimensional information expression capabilities, making them high-order features capable of fully expressing conditional information. The input process involves data format checks, dimension alignment, and numerical structure standardization to ensure that the aggregated features are compatible with the computational structure of the gated parameter projection layer, guaranteeing information integrity and computational accuracy during data transmission.

[0064] The gated parameter projection layer is a computational module with dynamic parameter generation capabilities. It typically contains multiple independent parameter mapping paths, information control structures, and dynamic combination mechanisms, enabling it to dynamically generate the parameter set required for subsequent structures based on the aggregated features of the input. This structure achieves full mining of input information and dynamic adjustment of parameter expression capabilities through multi-path parallel computation, dynamic information fusion, and nonlinear expression control. The parameter mapping paths in the gated parameter projection layer can be based on fully connected layers, convolutional structures, gated units, or other parameter mapping mechanisms, generating multiple sets of intermediate parameter expressions for different information expression dimensions and structural features.

[0065] The process of dynamically generating reversible transformation parameters for a flow model involves multi-path information transformation and dynamic combination of the input aggregated features through a gated parameter projection layer, forming a set of reversible transformation parameters with complete expressive power, standardized structure, and numerical accuracy. These reversible transformation parameters include scaling parameters and translation parameters, used for scaling and numerical translation operations on the input data within the flow model structure, respectively. The dynamic generation process relies on higher-order information representation in the aggregated features, ensuring that the generated parameters possess sensitive responsiveness and dynamic adjustment capabilities to input conditional information. The numerical structure and computational accuracy of the reversible transformation parameters directly affect the reversibility and numerical stability of the subsequent flow model, requiring strict parameter mapping and control mechanisms to guarantee their expressive effectiveness and numerical rationality.

[0066] In different implementation scenarios, aggregated features can be input to the gating parameter projection layer through memory transfer, data buffering mechanisms, parameter binding, or remote procedure calls. For local computing architectures, memory transfer and direct invocation can be used to input aggregated features into the gating parameter projection layer structure. For distributed deployments or heterogeneous system architectures, parameter synchronization mechanisms, data cache queues, or network transmission channels can be used to efficiently transfer aggregated features between different computing nodes, ensuring the integrity and real-time performance of data transmission.

[0067] In the design of the gated parameter projection layer structure, a dual-path structure can be used, setting independent scale parameter generation paths and translation parameter generation paths to perform feature mapping and parameter transformation on the input aggregated features respectively. Alternatively, a multi-path structure can be designed to capture the multi-dimensional expressive features of the input information through different paths, improving the expressive power of parameter generation and the efficiency of information utilization. The gating structure can employ gating units, dynamic weight adjustment, or attention control mechanisms to achieve dynamic combination of multi-path information and flexible adjustment of parameter expression.

[0068] In the process of dynamically generating reversible transformation parameters, the parameter mapping path can be based on multilayer perceptrons, convolutional neural networks, residual connection structures, or other deep representation mechanisms to enhance the nonlinear expressiveness and numerical stability of the generated parameters. The generated scaling parameters are typically ensured to meet the numerical constraints of subsequent reversible calculations through exponential mapping, positive value constraints, or other numerical restriction mechanisms. The generation process of translation parameters can incorporate the numerical distribution characteristics of the input information to improve the expressive flexibility and data adaptability of the translation operation.

[0069] Example Explanation: In the healthcare business domain, aggregated features can extract patient diagnostic information, medical record data, and real-time monitoring indicators to form high-order representations, which are then input into a gated parameter projection layer structure. The gated parameter projection layer dynamically generates scale and translation parameters based on different patient information, enabling the medical data flow model to dynamically adapt and adjust its representation to individual patient information, thereby improving the personalization level of medical auxiliary diagnosis, health risk prediction, and intervention strategy generation.

[0070] In the fintech business, aggregated features can integrate user identity information, transaction behavior patterns, and risk assessment indicators, and input them into the gating parameter projection layer. By dynamically generating reversible transformation parameters, the flow model can dynamically adjust the data representation structure and transformation parameters according to different user characteristics, enhancing the responsiveness and information expression capabilities of financial data generation, risk control, and fraud detection systems, and improving the personalized service capabilities and dynamic risk prevention and control level of financial business systems.

[0071] This embodiment, through the aforementioned operations, can dynamically generate reversible transformation parameters that meet the structural requirements of the flow model based on the high-order representation features of the input condition information, avoiding the limitations on model adaptability imposed by traditional static parameter design methods. The gated parameter projection layer enhances the expressive power and numerical control accuracy of parameter generation through multi-path information representation, dynamic combination, and nonlinear expression enhancement, significantly improving the system's adaptability and dynamic adjustment capabilities to diverse condition information. The dynamically generated reversible transformation parameters ensure the reversibility, numerical stability, and expressive flexibility of the flow model structure, improving the overall system's generation effect and information expression accuracy.

[0072] S40, sample from a preset distribution to generate initial latent variables, and use the initial latent variables as the current input data;

[0073] In this embodiment, the process of sampling from a preset distribution to generate initial latent variables refers to generating a set of data sequences that meet the distribution characteristics requirements based on a pre-defined probability distribution model using mathematical sampling methods. This data sequence serves as the input of latent variables during the system initialization phase and participates in subsequent data processing and feature transformation. The preset distribution is usually derived from standard probability models, such as normal distribution, uniform distribution, Laplace distribution, etc., but more complex multidimensional joint distributions or custom distribution structures can also be used depending on the specific business scenario. The selection of the preset distribution is based on the overall requirements of the system for data structure, numerical range, and statistical characteristics, ensuring that the generated initial latent variables meet the input requirements of downstream models in terms of numerical expression, structural integrity, and information diversity.

[0074] Sampling operations are typically based on pseudo-random number generation algorithms, Monte Carlo methods, or other numerical computation mechanisms to ensure the randomness of the sampling results, the consistency of the numerical distribution, and the integrity of the representation range. The sampling results constitute the initial latent variables, possessing multidimensional expressive power, structural regularity, and controllable statistical distribution characteristics, providing a stable and effective input foundation for subsequent data transformation, feature extraction, and information fusion. The initial latent variables can be single-dimensional data, matrix structures, tensor representations, or other forms of multidimensional data sets; the specific form is determined based on system design requirements and subsequent structural compatibility.

[0075] The operation of using initial latent variables as current input data involves data transfer, format standardization, and input structure adjustment. Current input data typically has specific dimensional structures, numerical ranges, and data type requirements. Operations such as format conversion, dimension mapping, and numerical normalization are necessary to ensure that the initial latent variables do not introduce information loss, numerical offsets, or structural disorder during transfer. As input variables in subsequent parts of the system, current input data participates in reversible linear transformations, partitioning operations, and coupled computations, directly affecting the stability and continuity of the data representation structure, information flow path, and the overall computational logic of the system.

[0076] In different implementation scenarios, the preset distribution can be a standard normal distribution, with a mean of zero and a variance of an identity matrix. This ensures that the initial latent variables have a zero mean, unit variance, and a standardized numerical structure, suitable for system design requirements that demand numerical symmetry and stability of expression. Alternatively, based on the characteristics of business data, a multidimensional Gaussian mixture distribution, beta distribution, or other probabilistic models suitable for expressing data heterogeneity and structural complexity can be set, improving the diversity and flexibility of the initial latent variables in relation to system input.

[0077] Sampling operations can generate random number sequences that satisfy preset distribution characteristics using a pseudo-random number generator combined with standard mathematical libraries or custom computation functions. For multidimensional data structures, the sampling process needs to combine matrix computation, tensor operations, and data parallelism techniques to improve generation efficiency and numerical accuracy. For high-dimensional data requirements, low-rank approximation, structure compression, or incremental generation techniques can be used to reduce storage overhead and computational burden, ensuring overall system performance stability.

[0078] When using initial latent variables as current input data, dimension mapping operations can be used to adjust the data structure, ensuring compatibility between latent variables and subsequent input layers in terms of the number and order of dimensions. For different system input specifications, data format conversion, type standardization, and numerical range mapping can be combined to improve the consistency and accuracy of input data, avoiding structural mismatches or numerical anomalies that could affect system stability.

[0079] Example Explanation: In the healthcare field, a pre-defined distribution can be constructed based on statistical data of patients' physiological indicators, building a multidimensional joint distribution or a standard Gaussian distribution model. The initial latent variables generated by sampling can simulate the diversity and structural patterns of patients' vital signs data. These latent variables are then used as current input data in information processing processes such as disease prediction, health monitoring, and risk assessment, improving the healthcare system's adaptability to individual patient differences and the accuracy of information expression.

[0080] In the fintech business, pre-defined distributions can construct multi-dimensional data distribution models based on user transaction behavior data, market risk indicators, or financial product attributes. The initial latent variables generated by sampling possess a good numerical expression structure and information diversity, providing a stable input data foundation for financial data generation, risk modeling, and fraud detection. By using the initial latent variables as the current input data, the financial system can dynamically adjust the data structure, optimize the information expression path, and enhance its adaptability to complex financial scenarios and risk control level.

[0081] Through the above operations, this embodiment enables the system to quickly and accurately generate initial latent variables that meet statistical requirements based on a preset probability distribution model, ensuring that the input data possesses good randomness, numerical normalization, and structural integrity. Dynamic sampling and input mapping enhance the system's adaptability to diverse data structures and the flexibility of input representation, avoiding the performance limitations imposed by fixed input designs. As the current input data, the initial latent variables provide a stable input foundation for subsequent data transformation and information processing, ensuring the continuity of the overall information flow, the consistency of data representation, and the stability of the system's computational structure, thereby improving the system's generation effect and expressive capabilities.

[0082] S50, perform an invertible linear transformation on the current input data to generate the transformed input data;

[0083] In this embodiment, performing a reversible linear transformation on the current input data refers to the process of adjusting the structure, recombining the values, or transforming the representation space of the current input data by employing matrix transformations, tensor mappings, or other mathematical linear transformations, while maintaining information integrity and the reversibility of the transformation. Reversible linear transformations typically include, but are not limited to, matrix multiplication, orthogonal transformations, identity determinant transformations, and reversible convolution operations. These transformations ensure that no information loss, numerical shift, or dimensional distortion is introduced during the transformation process, supporting subsequent reverse reconstruction and information restoration.

[0084] The current input data is a dataset formed by the system after the previous round of data transformation, latent variable generation, or coupling operations. It meets the requirements of structural standardization, numerical standardization, and complete expression. The reversible linear transformation, based on a pre-defined transformation matrix or a dynamically generated parameter set, performs linear reorganization of the data structure, spatial transformation of the expression path, and logical adjustment of the information arrangement order for this input data. This improves the data's expression efficiency, structural compatibility, and information richness in subsequent processing stages.

[0085] The transformed input data refers to the dataset that, after undergoing the aforementioned reversible linear transformation operation, possesses a new structure, new numerical expression, and a new spatial layout. This transformed data serves as the input data source for subsequent segmentation, scaling, and coupling operations. The transformed input data not only maintains the integrity and accuracy of the information but also optimizes the structural representation of the data through linear transformation, enhancing the system's computational efficiency and expressive flexibility.

[0086] In practical applications, reversible linear transformations can employ reversible one-dimensional convolution operations. By constructing a convolution kernel matrix with unit determinant constraints, information is ensured not to be lost during the transformation process, facilitating subsequent inverse operations to restore the data. Alternatively, linear transformation structures can be constructed based on orthogonal matrices, using mathematical methods such as QR decomposition, Householder reflection, or Givens rotation to generate transformation matrices with both invertibility and structural stability.

[0087] In high-dimensional data structures, tensor transformations and multi-channel matrix operations can be combined. Through dynamically generated sets of transformation parameters, dimension-wise linear recombination and spatial mapping can be implemented to ensure the integrity of data representation and structural reversibility during the transformation process. For specific business scenarios, different forms of matrix structures and transformation algorithms can be selected based on the dimensionality, numerical range, and structural characteristics of the input data to improve data representation efficiency and system computational performance.

[0088] The transformed input data needs to be combined with numerical normalization during the generation process to ensure the consistency and standardization of the transformation results in terms of numerical range, expression structure and data type, so as to avoid the impact of numerical overflow, structural misalignment or data type incompatibility on the stability and expression accuracy of subsequent calculation processes.

[0089] Example: In the healthcare business field, reversible linear transformations can be used to construct data transformation structures based on patient physiological monitoring data through orthogonal matrices. This allows for the representation, recombination, and spatial mapping of multidimensional vital signs data, improving the standardization of the data structure and the accuracy of its expression. It also enhances the adaptability of health monitoring systems to individual patient differences and improves their data representation capabilities.

[0090] In the fintech business, reversible linear transformations can combine financial transaction data, risk indicators, or credit scoring information, and based on dynamically generated matrix parameters, implement linear recombination and expression optimization of high-dimensional data, improve the expression efficiency of data structures and the accuracy of information fusion, and enhance the information expression capabilities and data processing efficiency of financial systems in data generation, risk assessment, and fraud detection.

[0091] Through the aforementioned operations, this embodiment enables the system to flexibly adjust the expression form and spatial layout of input data without losing information or disrupting the data structure, thereby enhancing the diversity of data expression, structural compatibility, and information expression efficiency. Dynamic and reversible linear transformations avoid the limitations imposed by fixed transformation structures on the system's expressive capabilities, enhancing its adaptability to different data structures and business scenarios. The transformed input data possesses a sound structural expression and numerical standardization, providing a high-quality, standardized data input foundation for subsequent data segmentation, scaling, and information coupling operations, ensuring the continuity of the overall information processing process, the standardization of the expression structure, and the stability of the computational logic.

[0092] S60, the transformed input data is divided into a first data part and a second data part;

[0093] In this embodiment, dividing the transformed input data into a first data part and a second data part means that, based on data structure partitioning rules and space allocation strategies, the input data after reversible linear transformation is divided into two independent data subsets according to a preset partitioning logic, which are named the first data part and the second data part, respectively, to ensure that the data partitioning process has structural regularity, clear boundaries and information integrity.

[0094] The transformed input data refers to the dataset formed after a preceding reversible linear transformation operation. This dataset has undergone spatial reorganization and expression optimization in terms of numerical representation, structural layout, and information arrangement, possessing a standardized multidimensional structure, continuous numerical distribution, and complete information expression capabilities. Data segmentation, based on the dimensional information, structural features, and expression requirements of this dataset, divides the overall data into two sub-datasets with independent structures and complete information through methods such as dimensional partitioning, channel partitioning, or sequence partitioning.

[0095] The first data portion typically serves as the reserved information set for subsequent coupling operations, carrying key expressive information and the foundation of the data structure. The second data portion serves as the target set for subsequent transformation operations, carrying data information requiring further scaling, translation, and expression adjustments. Through clearly structured and logically standardized data segmentation operations, it is ensured that the data partitioning process does not disrupt the overall information expression logic, does not introduce data structure misalignment or information loss, and improves the structural efficiency and expressive accuracy of subsequent information processing.

[0096] In practical implementation, the division boundaries and data scale between the first and second data parts can be flexibly adjusted based on the channel dimension, sequence dimension, or spatial dimension of the data, using methods such as proportional division, dynamic proportional division, or adaptive structural division. For example, for image data, a 50% proportional division can be implemented along the channel dimension or spatial dimension, with the first half of the transformed input data as the first data part and the second half as the second data part, ensuring a clear data division structure and standardized proportions.

[0097] For sequence-type data or high-dimensional tensor data, the segmentation ratio and boundary can be dynamically set according to the data representation requirements and structural characteristics, achieving data segmentation operations with strong structural adaptability and high representational integrity. For multi-channel, multi-dimensional mixed structure data, the data segmentation path and data subset structure can be dynamically adjusted based on channel-first or dimension-first segmentation strategies to ensure that the segmentation results conform to the system's representation logic and data structure specifications.

[0098] During the data segmentation process, it is necessary to verify the data partitioning boundaries, the structural integrity of the sub-data parts, and the continuity of information in real time to avoid problems such as boundary misalignment, data structure breakage, or missing information expression during the partitioning process, and to ensure that the first data part and the second data part have good structural independence and expression integrity.

[0099] Example: In the field of healthcare, for multi-channel physiological monitoring data, the transformed input data is divided into a first data part and a second data part by channel partitioning. The first data part is used to retain basic physiological information, while the second data part is used to implement dynamic transformation and expression optimization for target indicators, thereby improving the flexibility of health data expression and the adaptability of its structure.

[0100] In the field of fintech business, for multi-dimensional financial transaction data or risk indicator data, the transformed input data is divided into a first data part and a second data part by using sequence dimension partitioning and structure adaptive partitioning. The first data part retains the core transaction features, while the second data part implements dynamic scaling transformation and expression optimization, thereby improving the financial system's ability to express changes in transaction patterns and fluctuations in risk indicators and the efficiency of data structure processing.

[0101] Through the above operations, this embodiment enables the system to flexibly partition and separate the data set without disrupting the overall data structure and expression logic, forming a first data part and a second data part with independent structure and complete information. This standardized and clear data segmentation improves the efficiency and accuracy of subsequent scaling, translation, and information coupling operations, ensuring a clear structure, rigorous logic, and complete information in the data processing process. This enhances the system's information expression capabilities and data structure adaptability under different data structures and application scenarios.

[0102] S70, using the scale parameter and translation parameter in the reversible transformation parameters, perform scale transformation and translation transformation on the second data part to generate the transformed second data part;

[0103] In this embodiment, the scale parameter and translation parameter in the reversible transformation parameters are used to perform scale transformation and translation transformation on the second data part to generate the transformed second data part. Specifically, in the data processing process, combined with the previously dynamically generated reversible transformation parameters, a strict numerical transformation operation is performed on the divided second data part to complete the scale adjustment and spatial translation of the data, optimize the data expression structure, and improve the data adaptability and information expression ability.

[0104] The reversible transformation parameters originate from the preceding supernetwork generation module and are obtained through dynamic mapping and parameter decomposition. They include scaling parameters for controlling data scale changes and translation parameters for adjusting the spatial position of the data. The scaling parameters are generally high-dimensional vectors or matrices that match the structure of the second data part, and are used to control the enlargement, reduction, or dynamic scaling of the data through numerical operations. The translation parameters are generally bias data of the same dimension, used for overall translation or spatial position adjustment operations.

[0105] The second data component is a set of sub-data sets formed by structurally partitioning the transformed input data, possessing independent structure and the ability to express complete information. The scaling operation involves element-wise multiplication of the numerical content of the second data component with the scaling parameter, adjusting the relative scale of the numerical expression to achieve dynamic scaling, structural adaptation, or representational reconstruction of the data. The translation operation, based on the scaling operation, involves element-wise addition of the translation parameter to the second data component, completing the adjustment of the overall data position and optimization of the representation space.

[0106] The above transformation process usually includes an exponential operation step, which involves performing an exponential operation on the scale parameter to generate an exponential scale parameter. This ensures that the scaling transformation process has positive numerical definiteness and expression stability, avoids data distortion or structural abnormalities caused by negative or abnormal values, and guarantees the reversibility of the transformation result and the continuity of data expression.

[0107] In the specific implementation, the scaling parameter can be based on a dynamically generated high-dimensional parameter matrix. Combined with mathematical operations, an exponential function operation is performed element-wise on the second data part to form an exponential scaling parameter. This exponential scaling parameter is then multiplied element-wise with the second data part to complete the scaling transformation. The translation parameter can be directly added element-wise to the scaled data to generate the transformed second data part.

[0108] To accommodate different data structures and representation requirements, the computational modes of scaling and translation transformations can be adapted to matrix operations, tensor transformations, or batch processing, ensuring the efficiency and numerical stability of the transformation operations. For example, for image data, scaling and translation transformations can be combined with pixel-level operations and channel-level batch processing to ensure the continuity of image representation and the preservation of details; for sequence-type or high-dimensional tensor data, transformation operations can be processed hierarchically based on time, space, or structural dimensions to ensure the integrity of data representation and the reversibility of the structure.

[0109] During the transformation process, it is necessary to monitor the numerical range, expression stability, and structural matching of the scaling and translation parameters in real time to avoid transformation deviations caused by parameter anomalies or expression distortions. This ensures that the transformed second data part has a reasonable structure, stable values, and complete information, meeting the needs of subsequent data coupling and expression optimization.

[0110] Example Description: In the field of healthcare, for multi-channel physiological monitoring data, the system uses dynamically generated scaling and translation parameters to perform scaling and translation transformations on the second data part, realizing dynamic scaling and position adjustment of physiological signals, improving the clarity and structural adaptability of health monitoring data, and enhancing the system's ability to express and process individual differences, physiological fluctuations and data anomalies.

[0111] In the field of fintech business, for multi-dimensional financial transaction data or risk indicator data, the system dynamically adjusts the data expression scale and structural position of the second data part through scaling and translation transformations, thereby improving the accuracy and adaptability of the data in expressing market changes and risk fluctuations, and enhancing the response speed and data processing efficiency of the financial system to complex transaction patterns and dynamic market structures.

[0112] Through the above operations, this embodiment enables the system to efficiently and stably perform scaling and translation transformations on the second data portion, dynamically adjusting the data structure and expression scale to enhance the flexibility of data expression and the adaptability of the structure. Exponential transformation ensures the positive definiteness of the scaling transformation and the continuity of expression. Element-by-element transformation and spatial adjustment optimize the data structure layout and information distribution, enhancing the system's data processing capabilities and structural adaptability under different data structures and expression requirements.

[0113] S80, the first data portion is coupled with the transformed second data portion to generate new current input data;

[0114] In this embodiment, the first data part is coupled with the transformed second data part to generate new current input data. Specifically, in the data transformation process, the first data part formed by the previous segmentation operation and the second data part after scale transformation and translation transformation are recombined into a new data structure according to the requirements of structural consistency and numerical correspondence, which serves as the input source for the next stage of data processing.

[0115] The first data part originates from the structural division of the input data after the previous transformation. It typically retains some of the original data's expressive structure and information content, possessing independent numerical expression and structural information. The transformed second data part is a set of sub-data formed by scaling and translation transformations controlled by reversible transformation parameters based on the first data part. It possesses expressive capabilities after numerical adjustment and structural optimization.

[0116] The coupling operation refers to restoring the first data part and the transformed second data part to a unified data set with the same structure as before the transformation, based on the original data's partitioning dimensions, structural positions, and numerical correspondences, using methods such as data splicing, structural reorganization, or sequence merging. This operation not only achieves data structure reconstruction and information integration but also ensures the reversibility and continuity of the overall data representation.

[0117] The new current input data is the data structure after the coupling operation is completed. It has complete numerical expression, structural information and reversibility properties. It serves as the input basis for subsequent data processing, transformation and expression optimization, ensuring the stability and expression integrity of the overall data flow of the system.

[0118] In practical applications, the coupling operation between the first data part and the transformed second data part can be implemented based on various data structures. For example, for matrix data or tensor data, the coupling operation can use concatenation operations along a preset partitioning dimension to concatenate the two data parts into a complete data structure according to their dimensional positions; for sequence data or high-dimensional arrays, the coupling operation can rearrange the data order and generate new current input data through index recovery, sequence recombination, etc.

[0119] During the coupling process, it is necessary to verify in real time the structural dimensions, data scale and positional correspondence between the first data part and the transformed second data part to ensure that the data structure after splicing or recombination is consistent with the overall structure of the original input data, and to avoid data distortion or expression errors caused by dimension mismatch or structural abnormalities.

[0120] For different data types and application requirements, the coupling operation can be flexibly adapted to matrix concatenation, tensor expansion, sequence merging or structure mapping, ensuring the structural integrity, information continuity and reversibility requirements of data expression, and improving the system's processing capability and expression efficiency under different data structures and complex expression requirements.

[0121] Example Description: In the field of healthcare, for multidimensional physiological signal data, the system dynamically integrates physiological data from different sources and channels by coupling the first data part with the transformed second data part, restores the overall signal expression structure, improves the accuracy and structural stability of the health monitoring system in expressing multi-parameter data, and enhances the ability to track and analyze individual differences and changes in physiological state.

[0122] In the field of fintech business, for multi-source financial indicator data or dynamic risk characteristic data, the system integrates locally unchanged financial data with data parts that have been adjusted by parameter control through coupling operations, dynamically reconstructs the complete financial data structure, improves the system's ability to express complex transaction structures and market risk changes, and enhances the financial system's response speed and processing efficiency to dynamic market environments and multi-source data structures.

[0123] Through the aforementioned coupling operation, this embodiment enables the system to efficiently and accurately integrate the first data portion and the transformed second data portion into new current input data, restoring the overall data structure and expressive power, and ensuring the structural consistency and information integrity of the data after multiple transformations and segmentations. This operation not only optimizes the data representation structure and improves the system's reversibility and expressive continuity, but also provides stable and reliable data input for subsequent multi-layer data processing and transformation operations, enhancing the system's stability and adaptability in complex data representation and high-dimensional information processing.

[0124] S90, Repeatedly perform reversible linear transformation, segmentation, transformation and coupling operations on the new current input data until output data is generated.

[0125] In this embodiment, reversible linear transformation, segmentation, transformation, and coupling operations are repeatedly performed on the new current input data until output data is generated. Specifically, the previously generated new current input data is used as the starting point of the loop. According to a fixed logical order, reversible linear transformation, data structure segmentation, scaling and translation of sub-data parts, and coupling operation between the first data part and the transformed second data part are performed in sequence. Based on this loop structure, the above processing flow is continuously iterated. Finally, after meeting the set conditions or reaching the specified processing depth, the final result data is output.

[0126] The new current input data originates from the complete data structure formed in the previous round of coupling operations, possessing expressive characteristics consistent with the scale and structure of the original input data, ensuring data consistency and structural integrity throughout the loop operations. Reversible linear transformation operations are used within the loop to perform linear mappings or transformations on the current data structure, enhancing the flexibility and reversibility of the data representation space. The partitioning operation, based on fixed or dynamic partitioning logic, divides the transformed input data into a first data part and a second data part, facilitating differentiated data representation and parameter control.

[0127] The scaling and translation operations, based on dynamically generated reversible transformation parameters, dynamically adjust the data scale and numerical distribution of the second data part, achieving local optimization and structural adaptation of the data representation space. The coupling operation reintegrates the first data part and the transformed second data part into new current input data, restoring the overall data structure and ensuring the continuity and structural integrity of subsequent loop operations.

[0128] This loop structure repeatedly performs the above operations to progressively optimize the data representation structure, improving the overall data transformation capability and accuracy, until the output conditions are met, generating the final output data. The output data is the final data structure formed after the loop structure ends, possessing complete expressive power, optimized numerical distribution, and structural adaptability, serving as the final output of the system.

[0129] In practical applications, the number of iterations can be flexibly set according to system design requirements, data structure complexity, or target expression accuracy. This typically includes fixed-number iteration control or dynamic condition-triggered termination mechanisms. In fixed-number iteration mode, the system presets a threshold for the number of iterations and repeatedly performs reversible linear transformations, partitioning, transformations, and coupling operations according to a predetermined number of iterations, ensuring that the data structure outputs a stable and reliable final result after multiple optimizations.

[0130] In dynamic condition control mode, the system monitors the representational characteristics, numerical distribution, or optimization indicators of the data structure in real time. When the data representation reaches the set thresholds for structural integrity, numerical stability, or representational accuracy, the loop automatically terminates, and the final result data is output. This mode enhances the system's adaptability to complex data structures and dynamic representation requirements, and strengthens the flexibility and adaptability of the overall processing flow.

[0131] To cater to different data types and system requirements, the specific implementation methods, parameter settings, and structural adaptation logic of each operation step in the loop can be flexibly adjusted to ensure that the system can achieve efficient and accurate data processing and result output under different data scales, structural complexities, and expression requirements.

[0132] Example Description: In the field of healthcare, for complex physiological data with multiple channels and dimensions, the system dynamically optimizes the structural expression and numerical distribution of physiological signals by cyclically executing reversible linear transformations, segmentation, transformation and coupling operations. This enhances the health monitoring system's ability to express and adapt to complex physiological states, and improves the dynamic tracking effect and data expression stability of individual health data.

[0133] In the field of fintech business, for dynamic financial indicator data, risk factor data, or multi-source transaction structure data, the system gradually improves the accuracy of financial data structure expression and risk adaptability through multiple iterations of data transformation and optimization operations. This enhances the system's stability in expressing complex market environments, dynamic transaction structures, and volatile financial indicators, as well as the depth of data processing, thereby improving the financial system's risk monitoring, dynamic assessment, and intelligent decision-making capabilities.

[0134] Through the aforementioned loop structure, this embodiment enables the system to continuously optimize the data representation structure and numerical distribution during multi-level and multi-stage data processing, thereby improving the overall data expressiveness, structural adaptability, and reversibility. This loop operation not only enhances the system's processing depth and accuracy for complex data structures but also avoids structural limitations and expression errors caused by single transformations, ensuring that the final output data possesses stable and reliable expressive capabilities and improving the system's overall expressive flexibility and adaptability.

[0135] This invention relates to the field of artificial intelligence technology and can be applied to business scenarios such as fintech and healthcare. It discloses a data generation method, apparatus, device, and medium based on reversible transformation parameters, comprising: acquiring conditional information and encoding it to obtain a conditional vector; inputting the conditional vector into a multi-layer feature encoder within a hypernetwork generation module; extracting high-order features from the conditional vector to obtain aggregated features; dynamically generating reversible transformation parameters for a flow model based on the aggregated features; sampling initial latent variables from a preset distribution and using them as current input data; performing a reversible linear transformation on the current input data to generate transformed input data; segmenting the transformed input data into a first data part and a second data part; performing a scale transformation and a translation transformation on the second data part using the scale and translation parameters in the reversible transformation parameters to generate the transformed second data part; coupling the first data part and the transformed second data part to generate new current input data; and repeatedly performing the reversible linear transformation, segmentation, transformation, and coupling operations until output data is generated. This invention combines a supernetwork structure with a multi-layer reversible transformation process by dynamically generating reversible transformation parameters of a flow model based on conditional information. This enables flexible control for different conditional information, improves the diversity and adaptability of conditional data generation, avoids the limitations of relying on fixed structure design in traditional methods, and has higher generation flexibility and generalization performance.

[0136] In one embodiment, step S10 above includes:

[0137] S101, Receive the input condition information and determine the type of the condition information;

[0138] S102, when the type of the condition information is a discrete category label, the discrete category label is converted into a one-hot code, and the one-hot code is mapped to the embedding space through a learnable embedding matrix to obtain an embedding vector;

[0139] S103, when the type of the condition information is continuous text description, the continuous text description is projected onto the vector space by a multilayer perceptron to obtain a projection vector;

[0140] S104, the embedding vector or projection vector is used as a condition vector.

[0141] In this embodiment, condition information is acquired and encoded to obtain a condition vector. Specifically, for different types of condition information input according to system application requirements, data structure parsing, category determination and numerical expression conversion operations are performed to finally express the condition information in a vector form with fixed structure and numerical characteristics, which facilitates the structure adaptation and parameter mapping of subsequent data processing modules.

[0142] Conditional information originates from external inputs and encompasses various data types and structural expressions, primarily including discrete category labels and continuous text descriptions. Discrete category labels are typically used to express discrete attribute information with a clear classification structure or a finite set of states, and are widely applied in classification tasks, state indications, or conditional control scenarios. In the specific implementation, after receiving discrete category label data, the system first determines the data structure and category expression form of the information to ensure accurate identification of the data type and appropriate selection of the processing path.

[0143] When the conditional information is determined to be a discrete category label, the system maps the category information into a high-dimensional sparse vector with a fixed structure based on a one-hot encoding mechanism. One-hot encoding converts the discrete category label into a uniform and easily computed numerical expression by setting one to the corresponding category position and zero to other positions, ensuring structural compatibility and numerical operation stability in subsequent data processing. To enhance the system's expressive power and adaptability to category information, the one-hot encoding result is further input into a learnable embedding matrix. The embedding matrix dynamically adjusts the expression structure and numerical distribution of category information in the embedding space through parameter training, generating embedding vectors with high-dimensional feature expressive power and structural adaptability, thereby improving the expression effect and structural compatibility of conditional information in the system data stream.

[0144] When the conditional information is determined to be a continuous text description, the system performs feature extraction and expression compression operations on the text information based on the multilayer perceptron structure. As a neural network structure with nonlinear expression capabilities, the multilayer perceptron dynamically extracts high-order semantic features and structural information from the text information through multilayer parameter mapping and activation function processing. It maps the continuous text description into a vector expression with a fixed structure and uniform numerical distribution, forming a projection vector with stable expression structure and high-dimensional feature capabilities, ensuring the structural uniformity and expression stability of the system when processing continuous text description information.

[0145] Finally, based on the type of conditional information and the corresponding encoding result, the system selectively outputs either the embedding vector or the projection vector as the conditional vector. The conditional vector has a unified structural expression, a stable numerical distribution, and a high-dimensional feature expression capability, serving as input information for subsequent data processing. This enables the system to achieve structural compatibility, unified expression, and data adaptation for different types of conditional information, ensuring the continuity, stability, and accuracy of the overall data processing flow.

[0146] For example, the conditional coding formula:

[0147] c embed =E·onehot(c w )c embed Let R represent the conditional embedding vector, which encodes the conditional information; E represents the learnable embedding matrix with dimension R. K×dWhere K is the total number of conditional categories and d is the dimension of the embedding vector; onehot(c w ) indicates the condition category label c w The result of one-hot encoding, which is a method to map category information into a high-dimensional sparse vector.

[0148] Through the aforementioned operations, this embodiment enables the system to be compatible with multiple types of conditional information input. Efficient encoding and conversion paths are designed for discrete category labels and continuous text descriptions, ensuring the uniformity of structural representation and the stability of numerical processing for different data types. The introduction of embedding matrices and multilayer perceptrons enhances the expressive power and structural adaptability of conditional information in high-dimensional feature spaces, improving the system's processing depth and accuracy for complex conditional information. The overall process effectively optimizes the representation structure of multi-source heterogeneous conditional information in a unified data stream, ensuring the system's expression stability and data processing efficiency in complex application scenarios.

[0149] In one embodiment, step S20 above includes:

[0150] S201, Standardize the condition vector to generate a standardized condition vector;

[0151] S202, input the standardized conditional vector into the multi-layer feature encoder in the hypernetwork generation module;

[0152] S203, the input features are processed sequentially through multiple coding layers in the multi-layer feature encoder;

[0153] S204, In each coding layer, rotational position coding is applied to the input features of the current layer to generate position-enhanced features;

[0154] S205, Perform a multi-head self-attention operation on the location enhancement features to generate self-attention features;

[0155] S206, Perform a position feedforward operation on the self-attention features to generate the output features of the coding layer;

[0156] S207, the output features of the coding layer are used as the current layer input features of the next coding layer;

[0157] S208, weighted aggregation of the output features of all coding layers to generate aggregated features.

[0158] In this embodiment, the conditional vector is input into the multi-layer feature encoder in the hypernetwork generation module to extract high-order features of the conditional vector and obtain aggregated features. Specifically, it refers to performing hierarchical progressive processing on the input conditional vector based on a multi-level neural network structure. Through layer-by-layer feature transformation, structural enhancement and information integration, a feature set with rich expressive power and deep semantic information is extracted, and finally aggregated features with unified structure and enhanced expressive power are formed as input information for subsequent data processing and parameter generation.

[0159] First, the system standardizes the input condition vectors. The standardization operation adjusts the mean, variance, or other statistical characteristics of the input data to eliminate differences in numerical scale. This ensures that condition vectors from different sources and with different numerical distributions have a unified data structure and numerical distribution before entering the multi-layer feature encoder. This avoids feature extraction bias and model expression instability caused by input differences, and ensures the numerical stability and structural compatibility of the data in the multi-level feature encoding process.

[0160] The standardized conditional vectors are input into the multi-layer feature encoder within the hypernetwork generation module. This module combines a dynamic parameter generation mechanism with a fixed-structure feature encoding network to achieve adaptive parameter adjustment and structural expression optimization, enhancing the system's flexibility and expressive power when processing diverse input conditional information. The multi-layer feature encoder comprises multiple hierarchical structures. Through progressively layered feature transformations and information integration operations, the system gradually extracts high-order representational information and deep structural features from the input features.

[0161] Within each encoding layer, rotational position encoding is first applied to the input features of the current layer. Rotational position encoding is based on a differentiable position embedding strategy. By mapping the input features to a rotating coordinate system, it achieves efficient encoding and structural enhancement of positional information, improves the model's ability to express sequence structure and positional relationships, and ensures that the system has good positional representation ability and structural adaptability when processing input data with sequence features or structural dependencies.

[0162] The rotated position-encoded features are input into the multi-head self-attention structure as position-enhanced features. The self-attention mechanism captures long-distance dependencies and global associations in the input features by constructing dynamic weight relationships between features, thereby improving the system's ability to model complex feature relationships and deep representational structures. The multi-head structure enhances the system's representational dimensionality and information capture capability by parallelizing multiple independent attention subspaces, ensuring the diversity and stability of information representation in the high-dimensional feature space.

[0163] Features after the self-attention operation are further input into the positional feedforward network. The positional feedforward operation enhances the expressive complexity and structural nonlinearity of the features through nonlinear mapping and parameter transformation, improving the flexibility and expressive hierarchy of feature transformation, and ensuring the stability and structural integrity of feature representation in the multi-layer progressive process. The features output by each encoding layer serve as the current input to the next encoding layer. Through progressive feature enhancement and information integration, the system gradually constructs a high-order expressive structure and deep information representation of the input data.

[0164] The output features of all coding layers are ultimately integrated into a unified aggregated feature through a weighted aggregation operation. The aggregation operation dynamically adjusts the contribution of each layer's output features to the overall representation based on the weight parameters, ensuring that the system has flexible feature integration capabilities and representation adaptability under different application requirements. This generates aggregated features with complete structure, rich expressive power, and high-dimensional information carrying capacity, which serve as input information for subsequent dynamic parameter generation and data processing.

[0165] For example, the multi-layer feature results output by the feature encoder can be combined using a skip connection method to form aggregated features, specifically expressed as follows:

[0166]

[0167] Where h is the final aggregated feature, α i h represents the learnable weight coefficients corresponding to the output features of the i-th layer. i This represents the output of the feature encoder at layer i. This expression dynamically weights features from different layers during the overall fusion process, ensuring that the aggregated features can integrate multi-layer information and reflect global expressive power.

[0168] After obtaining the aggregated features, the parameters of each transformation layer of the flow model are further dynamically generated using the following expression:

[0169]

[0170] in, Let be the parameter set corresponding to the k-th transformation layer, W1, W2, and U be the learnable weight matrices, b1 and b2 be the learnable bias terms, h be the aggregated features calculated above, GeLU be the nonlinear activation function with smoothing properties, and ⊙ be the element-wise multiplication operation.

[0171] In this expression, the first part of the transformation result is generated by performing a linear transformation on matrix W1 and aggregated feature h, combined with a bias term b1, and applying the GeLU nonlinear function. Simultaneously, the aggregated feature h is linearly transformed with matrix U, and a bias term b2 is added to generate the second part of the transformation result. The two transformation results are fused using an element-wise multiplication structure to ensure efficient interaction of feature information. Finally, the fused result is input into matrix W2, and the output is the dynamic parameters of the invertible transformation layer.

[0172] This structure enables dynamic parameter generation based on conditional information. The parameter output results directly participate in the reversible mapping calculation of each transformation layer of the flow model, ensuring the flexibility and reversibility of the overall structure.

[0173] Through the above-described process, this embodiment enables the system to efficiently extract high-order features and deep representational structures from input conditional information, enhancing the expressive power and structural adaptability of conditional information in high-dimensional feature space. Standardization ensures the numerical stability of input data, rotational position encoding improves the system's positional information representation effect, multi-head self-attention mechanism enhances the global correlation expression ability between features, position feedforward operation improves the expressive complexity and hierarchical structure of feature transformation, and weighted aggregation mechanism optimizes the integration and expression of multi-layer feature information. The overall process improves the expression stability and data adaptability of conditional information during dynamic parameter generation, ensuring the system's expression accuracy and data processing efficiency in complex application scenarios.

[0174] In one embodiment, step S30 above includes:

[0175] S301, input the aggregated features into the gating parameter projection layer;

[0176] S302, the aggregated features are processed through the first linear transformation path of the gated parameter projection layer to generate the first transformation features;

[0177] S303, the aggregated features are processed through the second linear transformation path of the gated parameter projection layer to generate second transformation features;

[0178] S304, Perform a gating operation on the first transformation feature and the second transformation feature to generate a gating feature;

[0179] S305, The gated features are processed by linear transformation to generate a flow model parameter vector;

[0180] S306, the flow model parameter vector is divided into scale parameters and translation parameters to generate reversible transformation parameters.

[0181] In this embodiment, aggregated features are input to the gated parameter projection layer to dynamically generate reversible transformation parameters for the flow model. Specifically, the system uses a multi-path parameter mapping mechanism controlled by a gated structure to dynamically adjust the numerical expression of the reversible transformation parameters without altering the fixed structure of the flow model, achieving adaptive control of the flow model's transformation behavior under different conditions. This process, through multi-path feature transformation, gated control, and parameter decomposition operations, ensures that the generated parameters possess flexible expressive power, structural constraints, and numerical stability.

[0182] First, the system inputs the aggregated features into the gated parameter projection layer. The aggregated features, as a high-dimensional feature set containing conditional information, possess a complete conditional expression structure and high-order information carrying capacity. The gated parameter projection layer, as a mapping structure connecting the aggregated features and the parameter space of the flow model, achieves structural decoupling, expression enhancement, and parameter mapping of feature information based on multi-path linear transformation and gate control mechanisms.

[0183] The aggregated features are first input into the first and second linear transformation paths of the gated parameter projection layer, respectively. The two paths independently process the same aggregated features. The first linear transformation path generates the first transformed feature through operations such as matrix multiplication, bias adjustment, and nonlinear activation. The second linear transformation path generates the second transformed feature using the same or structurally variant mapping methods. Through dual-path transformation, the system introduces diverse representational structures into the parameter space, enhancing the flexibility and dimensionality of parameter mapping.

[0184] The first and second transformed features are further combined through a gating operation. This gating operation employs an element-wise dynamic weight control mechanism, adjusting the contribution of the two transformed features to the final representation based on the gating coefficients output by the activation function. This achieves dynamic pruning and information filtering of the feature representation. This operation effectively suppresses information redundancy and noise diffusion, ensuring that the generated features possess stable expressive power and structural compactness.

[0185] The gated features generated by the gating operation serve as input for the next step of parameter mapping. They are further processed through a linear transformation structure to generate the flow model parameter vector. The linear transformation, through matrix operations and bias adjustments, maps the gated features to the flow model parameter space, resulting in a parameter expression with structural integrity and numerical stability, satisfying the numerical and structural requirements of the downstream flow model's reversible transformation parameters.

[0186] Finally, the system divides the flow model parameter vector into scaling parameters and translation parameters according to a preset structure. The scaling parameters control the scaling ratio of each data dimension in the flow model, while the translation parameters control the numerical translation bias of each dimension. Together, they constitute a set of reversible transformation parameters. This set of parameters is used in the flow model to perform reversible affine transformations on the data at each layer, ensuring structural reversibility and numerical stability during data flow.

[0187] This embodiment implements a dynamic parameter generation process based on aggregated features through multi-path linear transformation, gating structure control, and parameter decomposition mechanisms. This process enhances the expressive flexibility of parameter mapping through a dual-path structure, improves the dynamic pruning capability of feature representation through gating operations, and ensures the stability and structural rationality of parameter numerical expression through linear mapping and parameter decomposition. Overall, it improves the system's dynamic adaptability to reversible transformation parameters of the flow model under different conditions, enhances the expressive effect of conditional control and the numerical stability of parameter mapping, meets the requirements of high-dimensional flow models for reversibility, stability, and dynamic expression of transformation parameters, and improves the overall flexibility and accuracy of the data generation process.

[0188] In one embodiment, step S40 above includes:

[0189] S401, Obtain preset distribution parameters including distribution type identifier and distribution characteristic parameters;

[0190] S402, parse the distribution type identifier to determine the probability distribution type;

[0191] S403, determine the distribution statistical characteristics based on the distribution characteristic parameters;

[0192] S404, Construct a probability distribution model based on the probability distribution type and distribution statistical characteristics;

[0193] S405, Generate a random number sequence that conforms to the probability distribution model;

[0194] S406, Normalize the random number sequence to generate a standardized random sequence as the initial latent variable;

[0195] S407, the initial hidden variable is used as the current input data.

[0196] In this embodiment, the system generates initial latent variables by sampling from a preset distribution and uses these latent variables as the current input data to construct the initial data representation basis for the conditional data generation process. This process, through the analysis of preset distribution parameters, construction of a probability model, random sampling, and data normalization operations, ensures that the generated latent variables possess structural integrity, standardized expression, and numerical stability, meeting the stringent requirements of the flow model input.

[0197] First, the system acquires preset distribution parameters, including a distribution type identifier and distribution characteristic parameters. The distribution type identifier is structural information used to distinguish the categories of probability models, and its sources may include identifiers for standard normal distribution, multivariate Gaussian distribution, uniform distribution, Laplace distribution, etc. The distribution characteristic parameters include a set of parameters that describe the specific numerical characteristics of the distribution. Common parameters include mean, variance, covariance matrix, range limits, etc. This set of parameters controls the numerical distribution shape and statistical properties of the generated data.

[0198] The system parses the distribution type identifier and determines the corresponding probability distribution type based on internal mapping relationships or an external configuration library. Different distribution types determine the mathematical model and sampling logic used in the data generation process. After determining the probability distribution type, the system calculates the distribution statistical characteristics based on the distribution characteristic parameters. The distribution statistical characteristics are used to quantitatively describe key indicators such as the central tendency, dispersion, and correlation structure of the data distribution, providing specific parameter support for the subsequent construction of the probability distribution model.

[0199] Based on the defined probability distribution type and calculated statistical characteristics, the system constructs a probability distribution model. This model is a mathematically encapsulated representation of the distribution, containing a complete definition of the probability density function or probability mass function, enabling numerical sampling and statistical analysis of high-dimensional data. After construction, the system generates random number sequences that conform to the probability distribution characteristics. The dimension of the random number sequence strictly matches the input dimension of the streaming model, ensuring data structure integrity and dimensional compatibility, and avoiding data structure anomalies or computational errors caused by dimensional mismatch.

[0200] The system performs normalization on the generated random number sequence. Normalization is used to standardize the numerical distribution of the data. It is usually achieved by means of linear transformation, mean adjustment and variance scaling to adjust the random number sequence to a standard range. Common target intervals include zero mean and unit variance or normalization to a fixed range. Normalization improves the stability of data representation, reduces the risk of extreme value interference, and optimizes the numerical adaptability and computational convergence efficiency of subsequent models.

[0201] After normalization, the system defines the standardized random sequence as the initial latent variable. This initial latent variable serves as input data for the conditional data generation process, carrying the structure of randomness and distribution control information, and possessing the initial expression function and random perturbation effect for data generation. The system further uses this initial latent variable as the current input data, feeding it into the subsequent reversible flow model to drive the continuous execution of the data generation process, ensuring the reversibility, stability, and expressive diversity of the data generation process.

[0202] This embodiment implements a data input initialization mechanism for stream models through distribution parameter analysis, probability model construction, random sampling, and normalization control. Based on flexible distribution type adaptation and numerical parameter configuration, this mechanism supports high-dimensional data generation under various probability models. The normalization operation further optimizes the stability and computational adaptability of data representation, comprehensively improving the structural integrity and accuracy of the data generation process. It enhances the system's flexibility in data generation under different control conditions and the numerical stability of stream model inputs, meeting the comprehensive requirements of high-dimensional conditional generation tasks for data diversity, structural stability, and random representation capabilities.

[0203] In one embodiment, step S60 above includes:

[0204] S601, determine the segmentation dimension of the transformed input data, and determine the total length of the segmentation dimension;

[0205] S602, the transformed input data is divided along the segmentation dimension according to the preset segmentation ratio to generate front-end data and back-end data;

[0206] S603, the front-end data is used as the first data part, and the back-end data is used as the second data part;

[0207] S604, verify whether the sum of the lengths of the first data portion and the second data portion in the segmentation dimension is equal to the total length;

[0208] S605, when equal to the total length, output the first data portion and the second data portion.

[0209] In this embodiment, the system performs structural segmentation on the transformed input data to construct local representation units and transformation control intervals, providing a structural foundation for subsequent scaling and translation transformations. This process, based on segmentation dimension determination, interval data partitioning, and structural integrity verification, ensures the logical integrity and accuracy of data expression during structural segmentation, meeting the structural requirements for multi-level data expression and conditional control.

[0210] First, the system parses the structural information of the transformed input data to determine the data segmentation dimensions. These dimensions are specific axes within the data structure and can originate from various representational levels, such as channel dimensions, spatial dimensions, and sequence dimensions. The system dynamically selects the most suitable segmentation dimension based on the flow model structure configuration or external parameter settings, ensuring the rationality and stability of the segmentation operation in terms of structural expression and logical partitioning. Once the segmentation dimensions are determined, the system calculates the total length of each dimension based on the data structure information. This total length serves as a complete scale indicator for the data within that dimension, providing a numerical basis for subsequent segmentation ratio calculations and interval data partitioning.

[0211] The system performs data structure partitioning operations along the partitioning dimension based on a preset partitioning ratio. The partitioning ratio can be flexibly set, commonly taking the form of equal-ratio partitioning, dynamic-ratio partitioning, or adaptive-ratio partitioning adjusted based on conditional information. The ratio parameter can originate from model configuration, external input, or dynamically generated logic. Based on the partitioning ratio, the system calculates the length range of the preceding and following data segments. Based on this range, the system extracts the preceding data segment as the first data part along the partitioning dimension. This first data part carries local representation information of the data structure and is typically used for direct transmission and structure preservation. Simultaneously, the system extracts the following data segment as the second data part. This second data part constitutes the main target region for transformation operations and serves as the effective range for subsequent scaling and translation transformations.

[0212] After the data structure is partitioned, the system performs a structural integrity check based on the total length of the partitioning dimension. This involves calculating the actual lengths of the first and second data parts along the partitioning dimension and verifying that the sum of their lengths is strictly equal to the total length. This check ensures the integrity of the data partitioning operation, avoiding inconsistencies in the data structure caused by numerical deviations, parameter errors, or data truncation. The system outputs the first and second data parts only if the check result satisfies the condition that the sum of their lengths equals the total length. This ensures the consistency and integrity of the data structure, improves the stability and reliability of the data partitioning process, builds a rigorous data structure foundation, and supports the smooth execution of subsequent conditional control and data transformation operations.

[0213] This embodiment achieves a highly stable structural segmentation mechanism for transformed input data by selecting segmentation dimensions, dividing structural intervals, and verifying data integrity. It supports flexible segmentation and structural organization of multidimensional data. The segmentation process effectively avoids expression deviations caused by structural inconsistencies or missing data through proportional control and structural verification. It ensures the integrity and controllability of local data structural expression, improves structural compatibility and expression accuracy in subsequent scaling, translation, and data coupling processes, and enhances the structural stability and multi-layer expression capability of the overall data generation process.

[0214] In one embodiment, step S70 above includes:

[0215] S701, Perform exponential operation on the scaling parameter in the reversible transformation parameters to generate exponential scaling parameters;

[0216] S702, Multiply the exponential scaling parameter element-wise with the second data portion to obtain scaled data;

[0217] S703, the translation parameter in the reversible transformation parameters is added element by element to the scaling data to obtain the transformed second data part.

[0218] In this embodiment, the system establishes a flexible numerical adjustment and representation reconstruction mechanism by performing scaling and translation transformations on the second data portion. This enables controllable enhancement and transformation of local features while maintaining the overall continuity of the data structure. This process relies on the scaling and translation parameters in the reversible transformation parameters, combined with exponential operations, element-wise operations, and data recombination operations, to construct a highly stable and expressive data transformation chain, thereby improving the numerical stability and expressive flexibility of the overall data generation process.

[0219] First, the system obtains the scaling parameter from the reversible transformation parameters. The scaling parameter is a numerical weight for controlling the local structure of the data, usually derived from the dynamic condition generation process or external control logic, and possesses adjustability and expressive adaptability. To ensure the numerical stability and monotonicity of the scaling transformation operation, the system performs exponential operation processing on the scaling parameter. Through the monotonically increasing and positive output characteristics of the exponential function, the scaling parameter is mapped to an exponential scaling parameter. The exponential scaling parameter is strictly greater than zero in the global range, avoiding numerical reversal, incorrect scaling direction, or discontinuity problems during the scaling transformation process, thus ensuring the numerical stability and expressive consistency of the transformation link.

[0220] The system performs element-wise multiplication of the generated exponential scaling parameters with the second data component. This element-wise operation executes numerical multiplication at corresponding positions within the data, achieving a numerical scaling process based on local scale control. This operation can flexibly adapt to different data structures and dimensions, ensuring the consistency and continuity of expression in the local structure of the transformation chain. Through element-wise multiplication, the system generates scaled data. This scaled data, in terms of numerical distribution, local representation, and overall structure, demonstrates condition-controlled variability and expressive adaptability, laying the foundation for enhanced expression and feature adjustment of the data at the structural level.

[0221] After generating the scaled data, the system obtains the translation parameters from the reversible transformation parameters. These translation parameters are numerical offset control parameters specific to the data structure, also derived from the dynamic condition generation process or external control logic, and possess flexible adjustment and representation reconstruction capabilities. The system performs element-wise addition of the translation parameters and the scaled data. Element-wise addition is performed at the corresponding positions in the data. Based on the local representation structure, the system completes the overall numerical offset and feature recombination process, improving the flexibility and diversity of data representation and enhancing the adaptability of conditional control to structural representation.

[0222] After completing the above operations, the system generates the transformed second data part. The transformed second data part demonstrates flexible control and expression reconstruction capabilities for the local structure of the data in terms of numerical scale, expression structure and condition adaptation. It provides a local data foundation with expression continuity, structural stability and multi-layer expression capabilities for subsequent data coupling and overall structure reconstruction, ensuring the structural integrity and expression consistency of the overall data generation process.

[0223] For example, the reversible transformation formula of the conditional flow generation module:

[0224] x2'=x2⊙exp(s c (x1))+t c (x1)

[0225]

[0226] x1 and x2 represent the first and second data parts of the input data; x2' represents the second data part after scaling and translation transformations; s c (x1) represents the scale parameters generated based on the first data portion, combined with the control information output by the hypernetwork; t c (x1) represents the translation parameters generated based on the first data portion, combined with the control information output by the hypernet; The control parameters output by the k-th layer supernetwork are split into scale parameters s after a segmentation operation. c Translation parameter t c ;⊙ represents element-wise multiplication; exp(·) represents exponential operation, ensuring that the scale parameter is strictly positive and meets the mathematical requirements of reversible transformation; Split(·) represents parameter splitting operation, dividing the generated control parameter into corresponding scale and translation parts.

[0227] To further improve the stability of the transformation and the numerical reliability of gradient calculation, the following regularization term is introduced in conjunction with the Jacobi regularization constraint:

[0228]

[0229] Where λ is the adjustment coefficient used to balance the ratio of the main loss to the regularization term, K is the total number of transformation layers, and J k Let ||·|| be the Jacobian matrix of the k-th transformation. F This indicates the calculation of the Frobenius norm, which reflects the overall numerical magnitude of the matrix.

[0230] This structure suppresses numerical divergence and gradient vanishing by constraining the overall numerical range of the Jacobian matrix of all transformation layers, thus ensuring the invertibility and stability of the overall data mapping.

[0231] In addition, to improve the generalization ability and cross-condition adaptability of parameter generation, a meta-learning strategy can be combined during the training phase, and the Reptile algorithm can be used to quickly iterate and update the hypernetwork parameters under multiple different conditions, which significantly improves the dynamic adaptation efficiency of reversible transformation parameters under new conditions.

[0232] In summary, parameter range constraints, Jacobi regularization, and meta-learning warm-up together construct a robust affine transformation process, ensuring the reversibility, numerical stability, and conditional adaptability of the second data part after scaling and translation transformations.

[0233] Example Explanation: In a personalized medical image generation scenario based on conditional information, the system dynamically generates high-quality, structured medical image data tailored to different patients' health information, examination categories, and diagnostic needs. This data is used to assist in disease risk prediction, personalized treatment plan development, and image data augmentation training. The specific operation process is as follows:

[0234] The system first acquires health condition information related to the target patient. This information comes from sources such as disease tags, treatment type identifiers, or structured text descriptions recorded in the electronic medical record system. For example, if a patient's condition information includes "chest CT scan" and "suspected nodule risk screening," the system determines the type of condition information. If it's a discrete category label, the system converts the category label into a numerical representation using one-hot encoding and maps this representation to an embedding space using a learnable embedding matrix to obtain an embedding vector. If the condition information is a continuous text description, the system performs a projection transformation on the text content using a multilayer perceptron network to obtain a projection vector. Finally, the system uses these embedding or projection vectors as condition vectors in the data processing flow.

[0235] The system inputs the conditional vector into the hypernetwork generation module, performing multi-layer feature encoding. First, the conditional vector is standardized to eliminate dimensional differences and numerical scale inconsistencies in the input data, ensuring numerical stability in subsequent multi-layer processing. Then, the standardized conditional vector is sequentially input into multiple encoding layers within the hypernetwork. Each encoding layer enhances its spatial structure representation through rotational position encoding, and then captures global dependency information through a multi-head self-attention mechanism. Further, it combines with a positional feedforward network to complete high-order feature extraction and representation reorganization. In all encoding layers, the system performs weighted fusion of the output features based on weight coefficients, ultimately generating aggregated features. These aggregated features comprehensively express patient characteristics, examination scenarios, and generation task requirements, providing a semantically rich and compact structural foundation for subsequent parameter control and data generation.

[0236] The system inputs aggregated features into the gating parameter projection layer, dynamically generating reversible transformation parameters for the flow model. Specifically, the aggregated features are processed through first and second linear transformation paths to form first and second transformed features. These two features are then dynamically fused through gating operations to form gated features with conditional adaptive expression. The system performs linear transformations on the gated features to generate flow model parameter vectors, which are further decomposed into scaling parameters and translation parameters. These two parameters together constitute the core control factors for data transformation and recombination at each layer of the flow model.

[0237] The system is based on a standard normal distribution and dynamically samples to generate initial latent variables as input information for the data stream. During the process, the system acquires preset distribution parameters, parses the distribution type identifier and distribution characteristic parameters, constructs a normal probability distribution model, and generates a random number sequence consistent with the input dimension of the stream model. To ensure numerical stability, the system performs normalization processing on the random number sequence to generate a standardized random sequence, which is then used as the initial latent variable input to the data stream transformation link.

[0238] The system first performs a reversible linear transformation on the current input data to generate transformed input data. Then, based on the set segmentation dimension and total length, it divides the transformed input data into a first data part and a second data part according to a preset ratio. In common cases, proportional segmentation can be performed based on the channel dimension to ensure data structure integrity and continuous representation. The system verifies the segmentation results to ensure that the first and second data parts strictly match the dimensions of the overall input data in terms of structural size, preventing structural damage caused by segmentation errors.

[0239] The system utilizes the scaling and translation parameters from the reversible transformation parameters to perform scaling and translation transformations on the second data part, improving the flexibility and condition adaptability of the local data representation. During the process, the system performs an exponential operation on the scaling parameter to generate an exponential scaling parameter, avoiding numerical inversion and expression anomalies. This parameter is then multiplied element-wise with the second data part to form scaled data. The translation parameter is further added element-wise with the scaled data, ultimately generating a transformed second data part with enhanced expression and stable structure.

[0240] The system couples the first data portion with the transformed second data portion, generating new current input data by splicing and recombining along the original segmentation dimension, ensuring data structure integrity and expression continuity. This process of reversible linear transformation, data segmentation, local transformation, and data coupling is iterated repeatedly. Based on the controlled number of transformation layers, the system repeats this data recombination link, constructing complex expression structures and multi-level transformation expressions layer by layer until the final medical image output data is generated.

[0241] For example, for suspected lung nodule screening tasks, the system can dynamically control the image generation process under different conditions and output diversified synthetic CT data with different lesion expressions, structural changes and imaging features. This can assist in clinical risk assessment and training with diverse data enhancement, thereby improving the quality of medical image generation, structural expression ability and application value.

[0242] In the fintech field, the system dynamically generates high-dimensional data representations with conditional constraints based on structured customer information, business scenarios, and risk control needs. These representations are applied to data augmentation and simulation scenarios such as credit approval, fraud detection, and compliance checks. The specific operation process is as follows:

[0243] First, the system acquires conditional information related to the target business object. This information can come from customer attribute tags, transaction category identifiers, or structured text descriptions. For example, for corporate lending, conditional information might include company type, industry affiliation, historical transaction patterns, and descriptions of fund usage. The system then determines the type of conditional information. If the information is a discrete category tag, such as company level, industry classification, or business category, the system uses one-hot encoding to convert the category information into a numerical representation. It then maps the encoding result to a unified vector space using a learnable embedding matrix, obtaining a compact and consistent embedding vector. If the information is a continuous text description, such as a description of fund usage or a customer risk statement, the system performs nonlinear projection on the text information using a multilayer perceptron structure to obtain a projection vector with both compression and feature preservation capabilities. Finally, the aforementioned embedding or projection vectors are used as conditional vectors in the data generation process.

[0244] The system inputs the conditional vectors into the hypernetwork generation module, sequentially performing multi-layer feature encoding operations. To improve the consistency of expression of conditional information from different sources, the system standardizes the conditional vectors, eliminating differences in scale and distribution of the input data and enhancing the adaptability of the downstream structure. The standardized conditional vectors are then input into the multi-layer feature encoder in the hypernetwork. Multiple encoding layers within the encoder process the input features sequentially. Each layer enhances the structural expressiveness of the input sequence through rotational position encoding, captures complex dependencies across dimensions and features using a multi-head self-attention mechanism, and improves the expression level and feature fusion effect through a positional feedforward structure. The output features of each layer are dynamically aggregated based on weight coefficients, ultimately generating aggregated features that comprehensively express customer background information, business risk attributes, and structural constraints of the generation task.

[0245] The system inputs aggregated features into the gated parameter projection layer to dynamically generate the reversible transformation parameters required within the flow model. Specifically, the aggregated features are transformed into two sets of transformation features through a first linear transformation path and a second linear transformation path, respectively. These two sets of transformation features are dynamically fused through a gating operation, enhancing the conditional adaptability and expressive flexibility of the parameter generation process. The fused gated features generate the flow model parameter vector through linear transformation. The system then divides the parameter vector into scaling parameters and translation parameters, providing a dynamic and controllable parameter basis for subsequent data flow transformation operations.

[0246] To construct the initial structure of the data stream, the system samples and generates initial latent variables from a preset distribution as the current input data. Specifically, this includes obtaining preset distribution parameters, parsing the distribution type identifier and distribution feature parameters, constructing a probability distribution model (such as a multidimensional normal distribution), generating a random number sequence that strictly matches the input dimension of the stream model, and performing normalization processing on the sequence to ensure that the initial latent variables have numerical stability and consistent expression, and participating in subsequent transformation processes as the current input data.

[0247] The system performs a reversible linear transformation on the current input data to generate the transformed input data. Based on the set segmentation dimension and total length, the system divides the transformed input data into a first data part and a second data part according to a preset ratio. The system verifies the segmentation results to ensure the dimensional accuracy and completeness of the structural division.

[0248] The system uses the generated scale and translation parameters to perform scale and translation transformations on the second data part in sequence. During the process, the system performs exponential operations on the scale parameters to generate exponential scale parameters, which are multiplied element-wise with the second data part to form scaled data. Then, the scaled data is added element-wise with the translation parameters to generate the transformed second data part, thereby enhancing the data's conditional adaptability and structural expressiveness.

[0249] The system concatenates and reassembles the first data part with the transformed second data part along the original segmentation dimension to generate new current input data, forming a complete structural representation. The above-mentioned reversible linear transformation, segmentation, local transformation, and coupled recombination process iterates repeatedly in the multi-layer structure. According to the number of transformation layers set by the control parameters, the system repeatedly performs dynamic recombination of the data structure and optimization of conditional representation until it finally generates high-dimensional output data that meets the conditional constraints.

[0250] In corporate credit simulation, the system can dynamically generate diverse synthetic data samples that conform to the distribution of business risks based on different corporate backgrounds and industry information. These samples can be applied to credit risk modeling, anti-fraud strategy verification, and data augmentation training, thereby improving the risk identification capability and business adaptability of financial technology systems under complex conditions. The system can also enhance the structural expressiveness and condition control capability of the data generation process, and meet the needs of financial scenarios for the reliability, controllability, and diversity of high-dimensional data.

[0251] This embodiment achieves a highly stable and expressive local data transformation mechanism based on reversible transformation parameters through exponential operations, element-wise scaling, and offset recombination. Exponential operations ensure positive output and numerical monotonicity in the scaling process, effectively avoiding numerical anomalies and discontinuous expression problems in the scaling process. Element-wise scaling and offset operations enhance the flexibility and multi-layer control capability of data expression, improve the adaptability of local structural expression and the effectiveness of conditional control, and comprehensively improve the stability of structural expression, the controllability of numerical recombination, and the adaptability to diverse expressions in the data generation chain.

[0252] In one embodiment, a data generation apparatus based on reversible transformation parameters is provided, which corresponds one-to-one with the data generation method based on reversible transformation parameters described in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the data generation device based on reversible transformation parameters of the present invention. The modules include a conditional encoding module 10, a feature extraction module 20, a parameter generation module 30, a latent variable initialization module 40, a reversible transformation module 50, a data segmentation module 60, a data transformation module 70, a data coupling module 80, and a multi-level iteration module 90. Detailed descriptions of each functional module are as follows:

[0253] The condition encoding module 10 is used to acquire condition information and encode the condition information to obtain a condition vector;

[0254] Feature extraction module 20 is used to input the conditional vector into the multilayer feature encoder in the hypernetwork generation module, extract the high-order features of the conditional vector, and obtain aggregate features;

[0255] The parameter generation module 30 is used to input the aggregated features into the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model;

[0256] The latent variable initialization module 40 is used to sample and generate initial latent variables from a preset distribution, and use the initial latent variables as the current input data;

[0257] The reversible transformation module 50 is used to perform a reversible linear transformation on the current input data to generate the transformed input data;

[0258] The data segmentation module 60 is used to segment the transformed input data into a first data part and a second data part;

[0259] Data transformation module 70 is used to perform scale transformation and translation transformation on the second data part using the scale parameter and translation parameter in the reversible transformation parameters to generate the transformed second data part;

[0260] The data coupling module 80 is used to couple the first data portion with the transformed second data portion to generate new current input data;

[0261] The multi-level iterative module 90 is used to repeatedly perform reversible linear transformation, segmentation, transformation and coupling operations on the new current input data until output data is generated.

[0262] In one embodiment, the conditional coding module 10 is specifically used for:

[0263] Receive input condition information and determine the type of the condition information;

[0264] When the type of the condition information is a discrete category label, the discrete category label is converted into a one-hot code, and the one-hot code is mapped to the embedding space through a learnable embedding matrix to obtain an embedding vector;

[0265] When the type of the conditional information is continuous text description, the continuous text description is projected onto the vector space by a multilayer perceptron to obtain a projection vector.

[0266] The embedding vector or projection vector is used as a condition vector.

[0267] In one embodiment, the feature extraction module 20 is specifically used for:

[0268] The condition vector is standardized to generate a standardized condition vector;

[0269] The standardized conditional vector is input into the multi-layer feature encoder within the hypernetwork generation module;

[0270] The input features are processed sequentially through multiple coding layers in the multi-layer feature encoder.

[0271] In each encoding layer, rotational position encoding is applied to the input features of the current layer to generate position-enhanced features;

[0272] Perform a multi-head self-attention operation on the location enhancement features to generate self-attention features;

[0273] Perform a position feedforward operation on the self-attention features to generate the output features of the coding layer;

[0274] The output features of the coding layer are used as the current input features of the next coding layer.

[0275] The output features of all coding layers are weighted and aggregated to generate aggregated features.

[0276] In one embodiment, the parameter generation module 30 is specifically used for:

[0277] The aggregated features are input into the gating parameter projection layer;

[0278] The aggregated features are processed through the first linear transformation path of the gated parameter projection layer to generate the first transformed features;

[0279] The aggregated features are processed through the second linear transformation path of the gated parameter projection layer to generate second transformed features;

[0280] Gating the first transformation feature and the second transformation feature together to generate a gated feature;

[0281] The gated features are processed by linear transformation to generate a flow model parameter vector;

[0282] The flow model parameter vector is segmented into scale parameters and translation parameters to generate reversible transformation parameters.

[0283] In one embodiment, the hidden variable initialization module 40 is specifically used for:

[0284] Obtain preset distribution parameters, including distribution type identifier and distribution characteristic parameters;

[0285] Parse the distribution type identifier to determine the probability distribution type;

[0286] The distribution statistical characteristics are determined based on the aforementioned distribution characteristic parameters;

[0287] Construct a probability distribution model based on the probability distribution type and distribution statistical characteristics;

[0288] Generate a random number sequence that conforms to the probability distribution model;

[0289] The random number sequence is normalized to generate a standardized random sequence as the initial latent variable;

[0290] Use the initial hidden variables as the current input data.

[0291] In one embodiment, the data segmentation module 60 is specifically used for:

[0292] Determine the segmentation dimension of the transformed input data, and determine the total length of the segmentation dimension;

[0293] The transformed input data is divided along the segmentation dimension according to a preset segmentation ratio to generate front-end data and back-end data.

[0294] The preceding data is taken as the first data part, and the following data is taken as the second data part;

[0295] Verify whether the sum of the lengths of the first data portion and the second data portion in the segmentation dimension is equal to the total length;

[0296] When the total length is equal to the given length, output the first data portion and the second data portion.

[0297] In one embodiment, the data transformation module 70 is specifically used for:

[0298] Perform exponential operation on the scaling parameter in the reversible transformation parameters to generate exponential scaling parameters;

[0299] The exponential scaling parameter is multiplied element-wise with the second data portion to obtain the scaled data;

[0300] The translation parameter in the reversible transformation parameters is added element-wise to the scaling data to obtain the transformed second data part.

[0301] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a data generation method based on reversible transformation parameters on the server side.

[0302] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the user-side functions or steps of a data generation method based on reversible transformation parameters.

[0303] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0304] Obtain condition information and encode the condition information to obtain a condition vector;

[0305] The conditional vector is input into the multi-layer feature encoder in the hypernetwork generation module to extract the high-order features of the conditional vector and obtain the aggregated features.

[0306] The aggregated features are input into the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model;

[0307] Initial latent variables are generated by sampling from a preset distribution, and these initial latent variables are used as the current input data;

[0308] Perform an invertible linear transformation on the current input data to generate the transformed input data;

[0309] The transformed input data is divided into a first data part and a second data part;

[0310] Using the scale parameter and translation parameter in the reversible transformation parameters, a scale transformation and a translation transformation are performed on the second data part to generate the transformed second data part;

[0311] The first data portion is coupled with the transformed second data portion to generate new current input data;

[0312] Repeatedly perform reversible linear transformation, segmentation, transformation and coupling operations on the new current input data until output data is generated.

[0313] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0314] Obtain condition information and encode the condition information to obtain a condition vector;

[0315] The conditional vector is input into the multi-layer feature encoder in the hypernetwork generation module to extract the high-order features of the conditional vector and obtain the aggregated features.

[0316] The aggregated features are input into the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model;

[0317] Initial latent variables are generated by sampling from a preset distribution, and these initial latent variables are used as the current input data;

[0318] Perform an invertible linear transformation on the current input data to generate the transformed input data;

[0319] The transformed input data is divided into a first data part and a second data part;

[0320] Using the scale parameter and translation parameter in the reversible transformation parameters, a scale transformation and a translation transformation are performed on the second data part to generate the transformed second data part;

[0321] The first data portion is coupled with the transformed second data portion to generate new current input data;

[0322] Repeatedly perform reversible linear transformation, segmentation, transformation and coupling operations on the new current input data until output data is generated.

[0323] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0324] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0325] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0326] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A data generation method based on reversible transformation parameters, characterized in that, Includes the following steps: Obtain condition information and encode the condition information to obtain a condition vector; The conditional vector is input into the multi-layer feature encoder in the hypernetwork generation module to extract the high-order features of the conditional vector and obtain the aggregated features. The aggregated features are input into the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model; Initial latent variables are generated by sampling from a preset distribution, and these initial latent variables are used as the current input data; Perform an invertible linear transformation on the current input data to generate the transformed input data; The transformed input data is divided into a first data part and a second data part; Using the scale parameter and translation parameter in the reversible transformation parameters, a scale transformation and a translation transformation are performed on the second data part to generate the transformed second data part; The first data portion is coupled with the transformed second data portion to generate new current input data; Repeatedly perform reversible linear transformation, segmentation, transformation and coupling operations on the new current input data until output data is generated.

2. The data generation method based on reversible transformation parameters as described in claim 1, characterized in that, Obtaining condition information and encoding the condition information to obtain a condition vector includes: Receive input condition information and determine the type of the condition information; When the type of the condition information is a discrete category label, the discrete category label is converted into a one-hot code, and the one-hot code is mapped to the embedding space through a learnable embedding matrix to obtain an embedding vector; When the type of the conditional information is continuous text description, the continuous text description is projected onto the vector space by a multilayer perceptron to obtain a projection vector. The embedding vector or projection vector is used as a condition vector.

3. The data generation method based on reversible transformation parameters as described in claim 1, characterized in that, The conditional vector is input into a multi-layer feature encoder within the hypernetwork generation module to extract higher-order features from the conditional vector, obtaining aggregated features, including: The condition vector is standardized to generate a standardized condition vector; The standardized conditional vector is input into the multi-layer feature encoder within the hypernetwork generation module; The input features are processed sequentially through multiple coding layers in the multi-layer feature encoder. In each encoding layer, rotational position encoding is applied to the input features of the current layer to generate position-enhanced features; Perform a multi-head self-attention operation on the location enhancement features to generate self-attention features; Perform a position feedforward operation on the self-attention features to generate the output features of the coding layer; The output features of the coding layer are used as the current input features of the next coding layer. The output features of all coding layers are weighted and aggregated to generate aggregated features.

4. The data generation method based on reversible transformation parameters as described in claim 1, characterized in that, The aggregated features are input into the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model, including: The aggregated features are input into the gating parameter projection layer; The aggregated features are processed through the first linear transformation path of the gated parameter projection layer to generate the first transformed features; The aggregated features are processed through the second linear transformation path of the gated parameter projection layer to generate second transformed features; Gating the first transformation feature and the second transformation feature together to generate a gated feature; The gated features are processed by linear transformation to generate a flow model parameter vector; The flow model parameter vector is segmented into scale parameters and translation parameters to generate reversible transformation parameters.

5. The data generation method based on reversible transformation parameters as described in claim 1, characterized in that, Initial latent variables are generated by sampling from a preset distribution, and these initial latent variables are used as the current input data, including: Obtain preset distribution parameters, including distribution type identifier and distribution characteristic parameters; Parse the distribution type identifier to determine the probability distribution type; The distribution statistical characteristics are determined based on the aforementioned distribution characteristic parameters; Construct a probability distribution model based on the probability distribution type and distribution statistical characteristics; Generate a random number sequence that conforms to the probability distribution model; The random number sequence is normalized to generate a standardized random sequence as the initial latent variable; Use the initial hidden variables as the current input data.

6. The data generation method based on reversible transformation parameters as described in claim 1, characterized in that, The transformed input data is divided into a first data part and a second data part, including: Determine the segmentation dimension of the transformed input data, and determine the total length of the segmentation dimension; The transformed input data is divided along the segmentation dimension according to a preset segmentation ratio to generate front-end data and back-end data. The preceding data is taken as the first data part, and the following data is taken as the second data part; Verify whether the sum of the lengths of the first data portion and the second data portion in the segmentation dimension is equal to the total length; When the total length is equal to the given length, output the first data portion and the second data portion.

7. The data generation method based on reversible transformation parameters as described in claim 1, characterized in that, Using the scale parameter and translation parameter in the reversible transformation parameters, a scale transformation and a translation transformation are performed on the second data portion to generate the transformed second data portion, including: Perform exponential operation on the scaling parameter in the reversible transformation parameters to generate exponential scaling parameters; The exponential scaling parameter is multiplied element-wise with the second data portion to obtain the scaled data; The translation parameter in the reversible transformation parameters is added element-wise to the scaling data to obtain the transformed second data part.

8. A data generation device based on reversible transformation parameters, characterized in that, The data generation device based on reversible transformation parameters includes: A condition encoding module is used to acquire condition information and encode the condition information to obtain a condition vector; The feature extraction module is used to input the conditional vector into the multi-layer feature encoder in the hypernetwork generation module, extract the high-order features of the conditional vector, and obtain aggregated features; The parameter generation module is used to input the aggregated features into the gated parameter projection layer to dynamically generate the reversible transformation parameters of the flow model; The latent variable initialization module is used to sample and generate initial latent variables from a preset distribution, and use the initial latent variables as the current input data; The reversible transformation module is used to perform a reversible linear transformation on the current input data to generate the transformed input data; The data segmentation module is used to segment the transformed input data into a first data part and a second data part. The data transformation module is used to perform scale transformation and translation transformation on the second data part using the scale parameter and translation parameter in the reversible transformation parameters to generate the transformed second data part; A data coupling module is used to couple the first data portion with the transformed second data portion to generate new current input data; The multi-level iterative module is used to repeatedly perform reversible linear transformation, segmentation, transformation and coupling operations on the new current input data until output data is generated.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a data generation program based on reversible transformation parameters stored in the memory and executable on the processor. When executed by the processor, the data generation program based on reversible transformation parameters implements the steps of the data generation method based on reversible transformation parameters as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a data generation program based on reversible transformation parameters, which, when executed by a processor, implements the steps of the data generation method based on reversible transformation parameters as described in any one of claims 1-7.

Images