Sequence data diffusion generation method for long dependency structure modeling
Patent Information
- Application Number
- CN202611030686.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-12
- Publication Date
- 2026-11-03
AI Technical Summary
Existing technologies for generating military situation sequence data suffer from problems such as difficulty in capturing long-range causal dependencies, insufficient multimodal feature fusion, inadequate privacy protection, and insufficient real-time generation capabilities, making it difficult to meet the requirements for generating high-quality, secure, and real-time military situation sequence data.
A three-stage progressive framework of causal enhancement representation, multimodal fusion diffusion generation, and privacy protection and real-time dynamic generation is adopted. Through Granger causality test, graph neural network, modality adaptive coding, causal guided denoising network and differential privacy training strategy, combined with streaming diffusion generation and state caching mechanism, high-quality, secure and real-time generation of military situation sequence data is achieved.
It significantly improves the causal consistency, structural fidelity, privacy security, and real-time response capabilities of military situation sequence data, supporting high-quality, high-security, and high-timeliness data augmentation for tasks such as operational situation awareness, action simulation, equipment support prediction, and communication link monitoring.
Smart Images
Figure CN123030947A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of artificial intelligence and sequence data generation, and specifically relates to a method for generating sequence data diffusion by modeling long dependency structures. Background Technology
[0002] With the rapid development of big data and artificial intelligence technologies in military intelligence construction, military situation sequence data plays an increasingly important role in tasks such as situational awareness, target activity pattern analysis, equipment support prediction, communication link monitoring, and action simulation. However, acquiring high-quality military situation sequence data faces many challenges in practical applications: on the one hand, raw military data often involves sensitive information such as mission deployment, target activities, equipment status, and communication links, making it difficult to directly share, publish, or use for large-scale model training due to security and privacy considerations; on the other hand, sequence data from real military scenarios suffers from problems such as scarce samples, insufficient samples of extreme events, class imbalance, and high annotation costs, severely restricting the effective application of data-driven methods such as deep learning in military situation simulation and decision support. Therefore, how to construct high-quality, controllable, and secure synthetic military situation sequence data through generation technology has become a key issue that urgently needs to be addressed in the field of military intelligent data processing.
[0003] Existing methods for generating sequence data can be mainly categorized as follows: Methods based on statistical models, such as Hidden Markov Models (HMMs) and Autoregressive Moving Average (ARMA) models, while possessing good interpretability, struggle to capture complex nonlinear dependencies and high-dimensional feature structures; methods based on Generative Adversarial Networks (GANs) can generate relatively realistic sequence samples, but suffer from problems such as training instability and mode collapse, particularly limiting their ability to model temporal dependencies in long sequences; methods based on Variational Autoencoders (VAEs) offer good generation diversity, but the fidelity and structural consistency of their generated samples still need improvement; and while diffusion models have made breakthrough progress in image generation in recent years, their application in generating structured sequence data is still in its early stages, with existing methods exhibiting significant shortcomings in long-range dependency capture, multimodal feature fusion, privacy protection, and real-time generation. Summary of the Invention
[0004] (a) Technical problems to be solved
[0005] The technical problem this invention aims to solve is how to provide a sequence data diffusion generation method for long-dependency structure modeling. By constructing a three-stage progressive framework of causal enhanced representation, multimodal fusion diffusion generation, and privacy protection and real-time dynamic generation, it focuses on addressing the following four core problems existing in the generation of military situation sequence data:
[0006] First, it addresses the problem of capturing long-range causal dependencies among military situation elements, breaking through the traditional method's reliance on short-term correlations, and realizing the identification and modeling of causal structures of variables such as target activities, equipment status, communication links, and combat events across long time scales.
[0007] Second, there is the issue of military multi-source heterogeneous sequence fusion, which unifies the processing of multiple modal features such as numerical sensor observations, categorical combat events, and textual mission records to achieve collaborative generation and semantic alignment of multimodal military information;
[0008] Thirdly, there is the issue of balancing privacy protection and utility of military sensitive data. Under strict security and privacy constraints, the statistical utility, situational evolution patterns, and causal consistency of the generated data must be maintained.
[0009] Fourth, it addresses the real-time generation problem in dynamic military scenarios, reduces the inference latency of diffusion models, and supports real-time response and incremental updates in situations of rapid changes in the situation.
[0010] (II) Technical Solution
[0011] To address the aforementioned technical problems, this invention proposes a sequence data diffusion generation method for long-dependency structure modeling. This method includes: a causal enhancement representation stage, a multimodal fusion diffusion generation stage, and a privacy protection and real-time dynamic generation stage.
[0012] In the causal augmentation representation stage, the input is the original multivariate situational sequence data. An improved Granger causality test method is used to obtain an initial causal graph. A graph neural network is used to model the causal graph as a directed weighted graph, and the edge strength is learned through a graph attention mechanism to obtain an optimized causal adjacency matrix. A sparsity loss function is introduced to remove redundant causal edges, resulting in an optimized situational causal graph. The original multivariate situational sequence and the causal structure are fused through a causal augmentation encoder to generate an augmented representation with causal semantics. Finally, the output is the optimized situational causal graph and the causal augmentation representation matrix.
[0013] In the multimodal fusion and diffusion generation stage, based on the causal enhancement representation matrix and the situation causal graph, a high-quality situation sequence sample is generated through the multimodal diffusion generation network, using modality adaptive coding, causal-guided denoising network, diffusion reconstruction loss, causal structure preservation loss, and multimodal alignment loss.
[0014] In the privacy protection and real-time dynamic generation phase, differential privacy constraints are applied to the training process involving sensitive information, and real-time response and incremental updates in dynamic scenarios are achieved through streaming diffusion generation and state caching mechanisms. A joint optimization objective of privacy-utility-latency is established, and the optimal privacy budget allocation and diffusion step configuration are adaptively selected.
[0015] (III) Beneficial Effects
[0016] This invention proposes a method for generating sequence data diffusion based on long-dependency structure modeling. By constructing a causal-enhanced long-dependency structure modeling module, a multimodal fusion diffusion generation framework, and a privacy-preserving and real-time dynamic generation mechanism, this invention achieves systematic improvements in four aspects: long-range dependency capture, multi-source heterogeneous feature fusion, sensitive data privacy protection, and real-time dynamic generation of military situation sequence data. This significantly enhances the causal consistency, structural fidelity, privacy security, and real-time response capability of generated military situation sequence data, effectively supporting the demand for high-quality, high-security, and high-timeliness sequence data augmentation in military missions such as combat situation awareness, action simulation, equipment support prediction, communication link monitoring, and simulation training. The main advantages are as follows:
[0017] (1) By combining Granger causality test with graph neural network, we can effectively identify causal dependencies across long time scales and significantly improve the causal structure fidelity of generated sequences compared with existing methods.
[0018] (2) A modal adaptive coding and causal guided denoising network were designed to realize the unified modeling and collaborative generation of multimodal features such as numerical, category, and text.
[0019] (3) Design a differential privacy training strategy for causal perception, allocate privacy budgets differently according to causal importance, and maximize the statistical utility and causal structure of the generated data while ensuring strict privacy constraints.
[0020] (4) Introducing a streaming diffusion generation and state caching mechanism reduces generation latency, supports online incremental updates in dynamic scenarios, and has good domain adaptability and real-time response capability. Attached Figure Description
[0021] Figure 1 This is a diagram illustrating the overall framework of the model of this invention. Detailed Implementation
[0022] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.
[0023] This invention relates to the field of military intelligent data generation and military situation sequence data modeling technology, specifically to a long-dependency structure modeling diffusion generation method for military situation sequence data. This method is applicable to the generation of structured sequence data in military application scenarios such as situation awareness, target trajectory analysis, equipment status monitoring, communication link status assessment, and action sequence simulation. The military situation sequence data includes, but is not limited to, target position and velocity sequences, radar or sensor observation sequences, equipment operational status sequences, communication network status sequences, combat event time sequences, and multi-source heterogeneous records during mission execution. This invention, by constructing a long-range causal dependency modeling mechanism and a multimodal fusion diffusion generation framework, achieves high-quality, high-fidelity, and highly adaptable military situation sequence data augmentation while ensuring military data security, privacy protection, and real-time response capabilities. This supports subsequent model training, situation simulation, simulation verification, and decision support tasks.
[0024] This invention proposes a method for generating sequence data diffusion based on long dependency structure modeling.
[0025] The framework diagram can be divided into three main parts: the causal augmentation representation stage, the multimodal fusion diffusion generation stage, and the privacy protection and real-time dynamic generation stage. Information flow and collaborative optimization between these stages are achieved through data structures such as causal graphs, augmented representations, and diffusion states.
[0026] In the causal augmentation representation stage, the input is the original multivariate situational sequence data. An improved Granger causality test method is used to obtain an initial causal graph. A graph neural network is used to model the causal graph as a directed weighted graph, and the edge strength is learned through a graph attention mechanism to obtain an optimized causal adjacency matrix. A sparsity loss function is introduced to remove redundant causal edges, resulting in an optimized situational causal graph. The original multivariate situational sequence and the causal structure are fused through a causal augmentation encoder to generate an augmented representation with causal semantics. Finally, the output is the optimized situational causal graph and the causal augmentation representation matrix.
[0027] In the multimodal fusion and diffusion generation stage, based on the causal enhancement representation matrix and the situation causal graph, a high-quality situation sequence sample is generated through a multimodal diffusion generation network, using modal adaptive coding, a causal-guided denoising network, diffusion reconstruction loss, causal structure preservation loss, and multimodal alignment loss. The sample is used for situation inference, simulation training, model training, and decision support.
[0028] In the privacy protection and real-time dynamic generation phase, differential privacy constraints are applied to the training process involving sensitive information, and real-time response and incremental updates in dynamic scenarios are achieved through streaming diffusion generation and state caching mechanisms. A joint optimization objective of privacy-utility-latency is established, and the optimal privacy budget allocation and diffusion step configuration are adaptively selected.
[0029] Example 1:
[0030] Overall framework diagram as follows Figure 1 As shown below, in conjunction with Figure 1 The specific implementation methods for each stage are described in detail.
[0031] Step 1: Causal Reinforcement Representation Stage
[0032] The core task of the causal enhancement representation stage is to perform structured modeling of the original multivariate sequence data of the military situation, identify the causal dependencies across time lengths between variables such as target state, sensor observation, equipment state, communication links, and combat events, and construct a causal enhancement military situation sequence representation to provide structural priors for the subsequent diffusion generation process.
[0033] Existing sequence modeling methods are mostly based on correlation or attention mechanisms, which struggle to distinguish between spurious associations and genuine causality, especially in long sequences where key dependencies are easily lost. To address this, this invention proposes a causal enhancement representation learning module that integrates Granger causality testing and graph neural networks. By combining statistical causal inference with structural learning, it extracts stable and interpretable long-range dependency structures from sequences.
[0034] Let the original military situation multivariate sequence dataset be... Where N is the number of samples, and each sample Indicates length is The variable dimension is A multivariate time series. The set of variables is defined as follows: This step aims to construct a directed causal graph. ,in Represents the set of causal edges between variables, and based on Generate enhanced sequence representations , This refers to the enhanced feature dimensions.
[0035] To identify long-range causal relationships between variables, this invention employs an improved Granger causality test method, introducing the maximum lag order. To capture causal effects over long time spans. For any two variables and Construct the following regression model:
[0036] (1)
[0037] in, For constant terms, and For regression coefficients, This is the error term. In the above model, Representing variables The influence of its historical values on the current value. Representing variables Historical values of variables The impact of the current value.
[0038] The null hypothesis of the F-test is , i.e., variables Historical sequence of variables It has no additional predictive power; the alternative hypothesis is that at least one exists. If the F-test rejects the null hypothesis, then it is considered that... yes Granger's reason.
[0039] To further enhance the robustness of causal discovery, a Bootstrap resampling method is introduced to resample the sample sequences multiple times and repeatedly calculate variable pairs. Test statistic or The value distribution is used to determine the causal significance threshold. In specific judgments, the threshold obtained from Bootstrap can be used together with a preset significance level for the determination. A causal edge is determined to exist between variables when the value is below the significance level or the Bootstrap threshold is met. Define the causal indicator variable:
[0040] (2)
[0041] in, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. For variables For variables Granger test value, The significance level is set to 0.05 by default. Therefore, when hour, , indicating from variable Pointer variable The causal edge exists; when hour, This indicates that the causal edge does not exist. The initial causal adjacency matrix is obtained by traversing all variable pairs. Construct an initial causal graph.
[0042] The initial causal graph may contain redundant edges or indirect dependencies, requiring further optimization. This invention employs a graph neural network to refine the causal structure, modeling the causal graph as a directed weighted graph, and learns the edge strength through a graph attention mechanism to obtain an optimized causal adjacency matrix.
[0043] Define node features as statistical embeddings of sequence variables The initial causal adjacency matrix serves as a priori for the graph structure, and the update formula for the graph convolutional layer is:
[0044] (3)
[0045] in, For variables In the set of neighbor nodes in a causal graph, For the first Edges in layered graph neural networks Attention coefficient For the first The learnable weight matrix of the layer, is the activation function. A layer refers to the first layer in a graph neural network. Sub-graph convolution or graph attention information propagation operations; as the number of layers increases, nodes can receive information from more distant causal neighbors layer by layer. Through multi-layer graph convolution, the representation of each node gradually incorporates information from its direct causal parent node and multi-hop causal ancestor nodes, thereby capturing complex dependencies on long-range causal chains.
[0046] To further remove redundant causal edges, a sparsity loss function is introduced:
[0047] (4)
[0048] in, This is the optimized causal adjacency matrix. Representing variables to variable Causal edge weights is the sparse regularization coefficient. Equation (4) applies a penalty to unnecessary edge weights by summing the absolute values of all edge weights; during the training and optimization process, if a causal edge contributes little to reducing the overall model loss, its edge weight will be gradually compressed to near 0. After training, the edge weights can be adjusted according to the preset edge weight threshold. The process involves filtering and removing edges below a threshold to eliminate redundant causal edges. The final output is a causal graph of the military situation. It retains significant long-range causal relationships and has structural sparsity, which facilitates the conditional guidance of subsequent diffusion models.
[0049] After obtaining the optimized military situation causality diagram Subsequently, this invention designs a causal augmentation encoder that fuses the original multivariate sequence of the military situation with a causal structure to generate an augmented representation with causal semantics. The causal structure refers to the optimized military situation causal graph. This represents the directed dependencies between variables, including the direction of causal edges, the weight of causal edges, and multi-hop causal paths formed by multiple causal edges. For example, in a military situation sequence, the target's motion state may affect sensor observations, the communication link state may affect command information transmission delays, and the equipment's operational state may affect mission execution capabilities. Specifically, for each time step... Define causal reinforcement features:
[0050] (5)
[0051] in Features of the original time step The graph neural network encoder outputs an augmented vector for each variable after fusing causal neighbors. The generated representation is obtained through equation (5). It not only contains observational information at the current moment, but also encodes long-range causal dependencies across variables, providing structured priors for subsequent diffusion generation.
[0052]
[0053] Step 2: Multimodal Fusion and Diffusion Stage
[0054] The core task of this step is to construct a multimodal conditional diffusion generation network based on the causal enhancement representation matrix generated in step one and the optimized military situation causal graph, so as to achieve high-quality generation of structured military situation sequence samples.
[0055] While existing diffusion models have achieved significant results in image generation, their direct application to structured sequence data faces three major challenges: first, it is difficult to uniformly model multimodal features (numerical, categorical, textual, etc.); second, the temporal dependency structure of sequences is easily destroyed during diffusion; and third, causal priors are difficult to effectively incorporate into diffusion conditions. To address these challenges, this invention proposes a Multimodal Conditional Diffusion for Sequences (MCDS) framework. Through modality-adaptive coding, a causal-guided denoising network, and three loss functions, it achieves high-quality, structure-preserving sequence generation.
[0056] Let the causal enhancement representation matrix output from step one be... The optimized military situation cause-effect diagram is as follows: .in, For sequence length, This refers to the feature dimension of the causal augmentation representation, which is the dimension obtained by concatenating the original military situation variable features with the causal neighborhood features output by the graph neural network. The goal of this step is to learn a multimodal fusion diffusion model. This generates new sequence samples that conform to the original military situation data distribution and maintain the causal structure.
[0057] To handle multimodal features in sequence data, this invention designs a Modality-Adaptive Encoding (MAE) module. Consider a real military situation sequence sample. Include There are three modalities, each with different data types and distribution characteristics. For numerical features, they are directly input after normalization; for categorical features, they are mapped to dense vectors using an embedding layer; for textual features, semantic embeddings are extracted using a pre-trained language model. The multimodal fusion encoding process can be represented as:
[0058] (6)
[0059] in, Indicates the actual sequence at time step The Modal features, For the first Dedicated encoders for each mode, For learnable modality fusion weights, For time step The multimodal fusion features. Through Equation (6), the model can adaptively fuse information from different modalities, avoiding training instability caused by differences in feature scale between modalities.
[0060] Causal-guided denoising networks are a core component of multimodal diffusion generation networks, responsible for predicting noise or directly predicting clean samples. To effectively integrate the causal structure information from step one into the generation process, this invention proposes a causal-guided Transformer denoising network (CGDT). This network uses noisy sequences... diffusion step Multimodal fusion features and cause-and-effect diagram Input: Predicted noise .
[0061] The core of CGDT is a causal-guided self-attention mechanism. Traditional Transformer attention mechanisms only calculate similarity based on sequence position, making it difficult to distinguish causal direction. This invention introduces a causal mask matrix. Its elements are defined as:
[0062] (7)
[0063] In self-attention computation, this mask restricts the flow of attention, ensuring that information propagates only along causal paths. The formula for causally guided self-attention computation is:
[0064] (8)
[0065] in, For element-wise multiplication, These are query, key, and value matrices, respectively. Let be the dimension of the key vector. In the specific calculation, first calculate... The attention score matrix is obtained, and its dimension is This represents the correlation strength between different time steps or variable positions in the sequence; then the causal mask matrix is used. Expand or map to the same dimension as the attention score matrix, allowing them to be multiplied element-wise. For Positions with a value of 0 have their attention scores suppressed; positions with a value of 1 retain their attention scores. Then... Normalization yields the attention weights after causal constraints, and they are then compared with... The output representation is obtained by multiplication. Through equation (8), the denoising network follows the causal graph at each denoising step. The constraints ensure the causal consistency of the generated sequences.
[0066] The aforementioned causal-guided denoising network is used to gradually remove noise during the backdiffusion process, progressively restoring random noise sequences into generated sequences that conform to the distribution of military situation data. To ensure that the generated results not only closely approximate real samples in numerical distribution but also maintain causal relationships and multimodal semantic consistency among military situation variables, this invention designs three loss functions during the training phase: diffusion reconstruction loss, causal structure preservation loss, and multimodal alignment loss. These loss functions collectively constrain the denoising network, enabling the model to learn the generation mapping from noisy sequences to high-quality military situation sequence samples.
[0067] The diffusion reconstruction loss is calculated using the noise prediction loss from standard DDPM:
[0068] (9)
[0069] in, For the spread and reconstruction losses; Indicates the diffusion step Real military situation sequence samples and random noise Seeking expectations; This represents a real sequence sample without added noise; This represents noise sampled from a standard Gaussian distribution; For the first The noise scheduling coefficient for each diffusion step. From step 1 to step 2 The cumulative noise retention factor of the step; Indicates the first The noisy sequence obtained by the step; Indicates by parameters The noise predicted by the controlled causal-guided denoising network; For causal enhancement representation matrix, This is the optimized cause-effect graph of the military situation. The loss is achieved by minimizing the actual noise. With prediction noise The differences between them enable the model to learn how to recover high-quality sequence samples from noisy sequences.
[0070] The causal structure preservation loss ensures that the generated sequence maintains the same dependencies as the original sequence on the causal graph. It is defined as follows, by comparing the difference in mutual information between the generated and original sequences on causal edges:
[0071] (10)
[0072] in, Preserve loss for causal structure; Cause-and-effect diagram of the optimized military situation The set of causal edges in; Representing variables With variables There is a causal dependency between them; and These represent the nth sequence in the generated sequence. The first variable and the second Time series of 1 variable; and They represent the nth and nth digits in the real sequence, respectively. The first variable and the second Time series of 1 variable; The mutual information estimate is used to measure the strength of the statistical dependency between two variable sequences. Equation (10) constrains the difference in mutual information between the generated sequence and the real sequence on the causal edge, so that the generated military situation sequence maintains a causal dependency structure similar to the real data.
[0073] The multimodal alignment loss ensures that the generated sequence remains consistent with the original sequence at the multimodal feature level, and is defined as:
[0074] (11)
[0075] in, For multimodal alignment loss; The sequence length; The number of modes; Indicates the first Various modalities, such as numerical sensor observations, categorical combat events, or text-based mission logs; Indicates the actual sequence at time step The Modal characteristics; Indicates the generated sequence at time step The Modal characteristics; Indicates the first The encoder corresponding to each mode; Let L2 be the squared norm. Equation (11) constrains the distance between generated samples and real samples in each modality coding space, so that the generated sequence maintains semantic consistency at the level of multi-source military information.
[0076] The total loss function is the weighted sum of the three:
[0077] (12)
[0078] in, The total loss function for model training; The diffusion reconstruction loss is used to constrain the accuracy of noise prediction. The loss is used to preserve the causal structure and constrain the consistency between the generated sequence and the real sequence in terms of causal dependencies. This is a multimodal alignment loss used to constrain the consistency between the generated sequence and the real sequence in different modal feature spaces; and To balance the hyperparameters, the weights of the causal structure preservation loss and the multimodal alignment loss in the total loss are adjusted respectively.
[0079] After training, the inference process for generating new sequences is as follows: First, initial noise is sampled from a standard Gaussian distribution. This noise can be understood as the first The initial state of the random sequence at each diffusion step; then following the reverse diffusion process from... Iterate step by step to Each step utilizes a causal guided denoising network (CGDT) to predict the noise components in the current noisy sequence, and combines this with a causal enhancement representation matrix. Cause-and-effect diagram of the optimized military situation Update the current sequence:
[0080] (13)
[0081] in, Indicates the first Noisy sequence states of one reverse diffusion step; Indicates by The noisy sequence state obtained from the previous time step after denoising; This represents the total number of diffusion steps. Number the current backdiffusion step; For the first The noise scheduling coefficient of the step; This is the cumulative noise retention factor; To improve the denoising network, conditions Z, multimodal fusion features E, and military situation causal graph are considered. Noise predicted under constraints; For the first The standard deviation of the random noise added in each step; It is standard Gaussian random noise. The meaning of this process is: the model first generates the first... The random noise sequence is processed step by step, guided by causal enhancement representation and military situation causal graph, to progressively denoise the sequence until a clean sequence at step 0 is obtained. This refers to the generated military situation sequence sample.
[0082] Step 3: Privacy Protection and Real-time Dynamic Generation Stage
[0083] The goal of the privacy protection and real-time dynamic generation phase is to achieve efficient and real-time generation of military situation sequence data while ensuring the privacy and security of sensitive military data. This meets the dual requirements of data security and response timeliness in military application scenarios such as combat situation awareness, equipment support prediction, communication link monitoring, and force action simulation. Existing sequence generation methods often employ simple anonymization or differential privacy-based noise enhancement for privacy protection, but this often leads to a significant decrease in generation quality. In terms of real-time performance, the iterative denoising process of traditional diffusion models has high computational overhead, making it difficult to support online generation requirements in dynamic scenarios. To address this, this invention proposes a real-time generation framework that integrates differential privacy protection and streaming diffusion generation mechanisms. Through adaptive allocation of privacy budget, lightweight inference paths, and incremental update strategies, it achieves an optimal balance between privacy protection and generation utility.
[0084] Let the multimodal fusion diffusion model trained in step two be... The goal of this step is to enable the real-time generation of new sequence samples using the model while protecting the privacy of the original training data, and to support online updates and rapid inference in dynamic scenarios.
[0085] To protect the privacy of training data from the source, this invention proposes a causal-aware differential privacy training strategy, utilizing the military situation causal graph generated in step one. The variables in the training samples and their corresponding model input channels, encoding channels, and related parameter gradients are grouped and pruned to reduce the utility loss caused by noise. It should be noted that this invention does not assume that the model parameters themselves have a causal relationship, but rather, based on the causal relationships between military situation sequence variables, model parameters or gradients corresponding to the same group of causally related variables are grouped into the same causal cluster. Specifically, based on the correlation strength of nodes in the causal graph, variables such as target state, sensor observation, equipment state, communication link state, and combat events are divided into several causal clusters, and the model parameter gradients processing these variables are assigned to the corresponding causal clusters. Parameters within each cluster share a privacy budget. For the first... For a causal cluster, the upper limit of sensitivity after gradient clipping is:
[0086] (14)
[0087] in, For the first Upper limit of gradient sensitivity for a causal cluster; and These are adjacent datasets, differing from each other by only one military situation sequence sample; All parameters of the model; In order to be with the first A subset of model parameters corresponding to each causal cluster; The model parameters are: The training dataset is The total loss calculated at that time; Represents the total loss relative to a subset of parameters. The gradient; Let L be the norm 2. Equation (14) represents the norm between any two adjacent datasets. Each causal cluster corresponds to the maximum magnitude of gradient change. By performing gradient pruning within causal clusters, the privacy-utility tradeoff can be controlled more precisely compared to global pruning.
[0088] Subsequently, Gaussian noise is added to the clipped gradient of each causal cluster to achieve... - Differential privacy protection:
[0089] (15)
[0090] in, For the first The gradient of a causal cluster after adding differential privacy noise; For the first The original gradients of the parameters corresponding to each causal cluster; This is used to clip gradients, ensuring that their L2 norm does not exceed the sensitivity upper bound. ; With a mean of 0 and a covariance of Gaussian noise; For noise scale, determined by privacy budget The size of the causal cluster is determined together with the size of the cluster. It is an identity matrix, and its dimension is the same as the first one. Each causal cluster corresponds to a gradient vector with the same dimension, which is used to represent the addition of mutually independent homoscedastic Gaussian noise to each gradient dimension. Through Equation (15), each causal cluster is assigned an independent amount of noise. The more tightly packed the causal structure of the cluster, i.e. the critical dependency path, the smaller the noise can be assigned, thereby protecting the generation quality of the critical causal link.
[0091] The allocation of privacy budgets follows the principle of causal importance: ,in For the variables in formula (1) right Granger regression coefficients, If the lag order is the largest, then the causal cluster Privacy Budget Positively correlated with the average importance of edges within the cluster:
[0092] (16)
[0093] in, For the total privacy budget, To be assigned to the Privacy budget for a causal cluster For the first The set of causal edges contained in a causal cluster Cause-and-effect diagram of the optimized military situation The set of all causal edges in the array. Causal edge The importance weight. It should be noted that... Indicates privacy budget, The two terms, "edge set" and "edge set," have different meanings. Equation (16) calculates the total privacy budget based on the proportion of edge weights within the causal cluster to the total edge weights. The critical causal paths are assigned to different causal clusters. Through adaptive allocation, the critical causal paths achieve a more reasonable privacy budget configuration, thereby striking a balance between privacy protection and generation quality.
[0094] In dynamic military scenarios, military situation sequence data may continuously arrive, such as radar detection sequences, target trajectory update sequences, equipment status monitoring sequences, or communication link status sequences. Traditional methods require regenerating data for each new sample, resulting in high computational overhead. This invention designs a state caching mechanism to cache and reuse the reverse process states of the diffusion model.
[0095] Let the current actual time be... The generation state is That is, at actual time The intermediate representations of each diffusion step in the cache. Among them, This refers to the total number of diffusion steps or the total number of denoising steps in the diffusion model. Indicates the first One diffusion step; Indicates actual time Next The intermediate sequence states corresponding to each diffusion step. When new data... Upon arrival, only partial updates are needed based on the existing state:
[0096] (17)
[0097] in, For the first The update step size coefficient for each diffusion step. For lightweight network updates, Indicates actual time Newly arrived incremental data on the military situation. This is the optimized cause-and-effect graph of the military situation. This lightweight update network is based on... To determine the variables and causal paths affected by new data, only the locally diffused states affected by the new data are updated. Through state caching and incremental updates, the computational cost of generating new samples is reduced from... Down to ,in The number of affected causal edges is typically much smaller than the total number of edges. .
[0098] There is an inherent trade-off between privacy protection and real-time generation: stronger privacy protection (less...) This requires more noise, which may lead to a decrease in generation quality, thus requiring more diffusion steps to recover quality and increasing latency. This invention establishes a joint privacy-utility-latency optimization objective, adaptively selecting the optimal privacy budget allocation and diffusion step configuration.
[0099] Define generation utility To generate the similarity between the sample and the original data distribution (e.g., the inverse of the maximum mean difference), delay Privacy costs for average generation time This represents the total privacy budget expenditure. The optimization objective is to maximize the weighted overall score.
[0100] (18)
[0101] in, The evaluation objectives are based on a comprehensive assessment of privacy, utility, and delay. This represents the privacy budget allocation scheme for each causal cluster; This represents the total number of denoising steps in the diffusion model; The generation utility, given a privacy budget allocation and number of diffusion steps, can be measured by the similarity between the generated samples and the distribution of real military situation data. This represents the generation delay, which typically varies with the number of diffusion steps. Increase as it grows; This indicates penalties related to the strength of privacy protection, when the privacy budget... When the value is smaller, privacy protection is stronger, but model training noise is greater. This term is used to characterize the impact of privacy protection on generative utility. and These are the delay penalty coefficient and the privacy penalty coefficient, respectively. Equation (18) means that, under the premise of meeting the privacy protection requirements of military sensitive data, the privacy budget of each causal cluster is jointly selected. and diffusion steps The goal is to maximize generation utility while controlling generation latency and quality loss due to privacy protection. This optimization problem can be solved using Bayesian optimization, which automatically selects the optimal privacy budget allocation strategy and diffusion steps based on the specific military application scenario.
[0102] Beneficial effects:
[0103] This invention achieves systematic improvements in four aspects of military situation sequence data: long-range dependency capture, multi-source heterogeneous feature fusion, sensitive data privacy protection, and real-time dynamic generation. These improvements are achieved by constructing a causal-enhanced long-dependency structure modeling module, a multimodal fusion and diffusion generation framework, and a privacy protection and real-time dynamic generation mechanism. This significantly enhances the causal consistency, structural fidelity, privacy security, and real-time response capabilities of generated military situation sequence data. It effectively supports the demand for high-quality, high-security, and high-timeliness sequence data in military missions such as operational situation awareness, force action simulation, equipment support prediction, communication link monitoring, and simulation training. The main advantages are as follows:
[0104] (1) By combining Granger causality test with graph neural network, we can effectively identify causal dependencies across long time scales and significantly improve the causal structure fidelity of generated sequences compared with existing methods.
[0105] (2) A modal adaptive coding and causal guided denoising network were designed to realize the unified modeling and collaborative generation of multimodal features such as numerical, category, and text.
[0106] (3) Design a differential privacy training strategy for causal perception, allocate privacy budgets differently according to causal importance, and maximize the statistical utility and causal structure of the generated data while ensuring strict privacy constraints.
[0107] (4) Introducing a streaming diffusion generation and state caching mechanism reduces generation latency, supports online incremental updates in dynamic scenarios, and has good domain adaptability and real-time response capability.
[0108] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for generating sequence data diffusion by modeling long-dependency structures, characterized in that, The method includes: a causal enhancement representation stage, a multimodal fusion diffusion generation stage, and a privacy protection and real-time dynamic generation stage; In the causal augmentation representation stage, the input is the original multivariate situational sequence data. An improved Granger causality test method is used to obtain an initial causal graph. A graph neural network is used to model the causal graph as a directed weighted graph, and the edge strength is learned through a graph attention mechanism to obtain an optimized causal adjacency matrix. A sparsity loss function is introduced to remove redundant causal edges, resulting in an optimized situational causal graph. The original multivariate situational sequence and the causal structure are fused through a causal augmentation encoder to generate an augmented representation with causal semantics. Finally, the output is the optimized situational causal graph and the causal augmentation representation matrix. In the multimodal fusion and diffusion generation stage, based on the causal enhancement representation matrix and the situation causal graph, a high-quality situation sequence sample is generated through the multimodal diffusion generation network, using modality adaptive coding, causal-guided denoising network, diffusion reconstruction loss, causal structure preservation loss, and multimodal alignment loss. In the privacy protection and real-time dynamic generation phase, differential privacy constraints are applied to the training process involving sensitive information, and real-time response and incremental updates in dynamic scenarios are achieved through streaming diffusion generation and state caching mechanisms. A joint optimization objective of privacy-utility-latency is established, and the optimal privacy budget allocation and diffusion step configuration are adaptively selected.
2. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 1, characterized in that, In the causal enhancement representation stage, an improved Granger causality test method is used to obtain the initial causal graph, including: Let the original situation multivariate sequence dataset be... Where N is the number of samples, and each sample Indicates length is The variable dimension is Multivariate time series; the variable set is defined as The goal is to construct a directed causal graph. ,in Represents the set of causal edges between variables, and based on Generate enhanced sequence representations , The enhanced feature dimensions; To identify long-range causal relationships among variables, an improved Granger causality test method is adopted, which introduces the maximum lag order. To capture causal effects over long time spans; for any two variables and Construct the following regression model: (1) in, For constant terms, and For regression coefficients, This is the error term; in the above model, Representing variables The influence of its historical values on the current value. Representing variables Historical values of variables The impact of the current value; The null hypothesis of the F-test is , i.e., variables Historical sequence of variables It has no additional predictive power; the alternative hypothesis is that at least one exists. If the F-test rejects the null hypothesis, then it is considered that... yes Granger's reason; The Bootstrap resampling method is introduced to resample the sample sequence multiple times, repeatedly calculating variable pairs. Test statistic or The value distribution is used to determine the causal significance threshold. In specific judgments, the threshold obtained from Bootstrap and the preset significance level are used together for the judgment. When the value is less than the significance level or the Bootstrap threshold is met, a causal edge is determined to exist between the variables; For the preset significance level, define the causal indicator variable: (2) in, This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. For variables For variables Granger test value, At the significance level; therefore, when hour, , indicating from variable Pointer variable The causal edge exists; when hour, This indicates that the causal edge does not exist; by traversing all variable pairs, the initial causal adjacency matrix is obtained. Construct an initial causal graph.
3. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 2, characterized in that, In the causal enhancement representation stage, a graph neural network is used to model the causal graph as a directed weighted graph, and the edge strength is learned through a graph attention mechanism to obtain the optimized causal adjacency matrix, including: Define node features as statistical embeddings of sequence variables The initial causal adjacency matrix serves as a priori for the graph structure, and the update formula for the graph convolutional layer is: (3) in, For variables In the set of neighbor nodes in a causal graph, For the first Edges in layered graph neural networks Attention coefficient For the first The learnable weight matrix of the layer, For activation function; the first A layer refers to the first layer in a graph neural network. Sub-graph convolution or graph attention information propagation operation; as the number of layers increases, nodes can receive information from more distant causal neighbors layer by layer; through multi-layer graph convolution, the representation of each node gradually integrates information from its direct causal parent node and multi-hop causal ancestor node, thereby capturing complex dependencies on long-range causal chains.
4. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 3, characterized in that, In the causal augmentation representation stage, a sparsity loss function is introduced to remove redundant causal edges, resulting in an optimized situational causal graph. A causal augmentation encoder then fuses the original multivariate situational sequence with the causal structure to generate an augmented representation with causal semantics. The final output consists of the optimized situational causal graph and the causal augmentation representation matrix, including: To further remove redundant causal edges, a sparsity loss function is introduced: (4) in, This is the optimized causal adjacency matrix. Representing variables to variable Causal edge weights For sparse regularization coefficients; Equation (4) applies a penalty to unnecessary edge weights by summing the absolute values of all edge weights; during the training optimization process, if a causal edge contributes little to reducing the overall model loss, its edge weight will be gradually compressed to near 0; after training, the edge weight is adjusted according to the preset edge weight threshold. The process involves filtering and removing edges below a threshold to eliminate redundant causal edges; the final output is a situational causal graph. It retains significant long-range causal relationships while also exhibiting structural sparsity; After obtaining the optimized situational cause-effect diagram Subsequently, a causal augmentation encoder was designed to fuse the original multivariate situation sequence with the causal structure, generating an augmented representation with causal semantics; the causal structure refers to the representation derived from the optimized situation causal graph. The directed dependencies between variables are represented, including the direction of causal edges, the weight of causal edges, and multi-hop causal paths formed by multiple causal edges; for each time step Define causal reinforcement features: (5) in Features of the original time step For the graph neural network encoder, the output is the augmented vector after fusing causal neighbors for each variable; the generated representation is obtained through equation (5). It not only contains observational information at the current moment, but also encodes long-range causal dependencies across variables, providing structured priors for subsequent diffusion generation.
5. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 4, characterized in that, In the multimodal fusion diffusion generation stage, modal adaptive coding includes: Let the causal enhancement representation matrix be... The optimized situational cause-effect diagram is as follows: ;in, For sequence length, The feature dimension for causal augmentation representation is the dimension obtained by concatenating the original situation variable features with the causal neighborhood features output by the graph neural network; the goal is to learn a multimodal fusion diffusion model. Generate new sequence samples that conform to the original situational data distribution and maintain the causal structure; To handle multimodal features in sequence data, a modality adaptive coding module (MAE) is designed; assuming real-world sequence samples... Include There are three modalities, each with different data types and distribution characteristics. For numerical features, they are directly input after normalization. For categorical features, they are mapped to dense vectors using an embedding layer. For textual features, semantic embeddings are extracted using a pre-trained language model. The multimodal fusion encoding process is represented as follows: (6) in, Indicates the actual sequence at time step The Modal features, For the first Dedicated encoders for each mode, For learnable modality fusion weights, For time step The multimodal fusion features; through equation (6), the model can adaptively fuse information from different modalities.
6. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 5, characterized in that, In the multimodal fusion diffusion generation stage, the causal-guided denoising network includes: A causal-guided denoising network is responsible for predicting noise or directly predicting clean samples. To effectively integrate causal structural information into the generation process, a causal-guided Transformer denoising network, CGDT, is proposed. This network uses noisy sequences... diffusion step Multimodal fusion features and cause-and-effect diagram Input: Predicted noise ; The core of CGDT is a causal-guided self-attention mechanism; it introduces a causal mask matrix. Its elements are defined as: (7) In self-attention computation, this mask restricts the flow of attention, ensuring that information propagates only along the causal path; the formula for causally guided self-attention computation is: (8) in, For element-wise multiplication, These are query, key, and value matrices, respectively. Let be the dimension of the key vector; in the actual calculation, first calculate... The attention score matrix is obtained, and its dimension is This represents the correlation strength between different time steps or variable positions in the sequence; then the causal mask matrix is used. Expand or map to the same dimension as the attention score matrix so that they can be multiplied element-wise; for Positions with a value of 0 have their attention scores suppressed; positions with a value of 1 retain their attention scores; then... Normalization yields the attention weights after causal constraints, and they are then compared with... The output representation is obtained by multiplication; through equation (8), the denoising network follows the causal graph in each denoising step. The constraints ensure the causal consistency of the generated sequences.
7. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 6, characterized in that, In the multimodal fusion diffusion generation stage, diffusion reconstruction loss, causal structure preservation loss, and multimodal alignment loss are used to generate high-quality situation sequence samples, including: The diffusion reconstruction loss is calculated using the noise prediction loss from standard DDPM: (9) in, For the spread and reconstruction losses; Indicates the diffusion step Real situation sequence samples and random noise Seeking expectations; This represents a real sequence sample without added noise; This represents noise sampled from a standard Gaussian distribution; For the first The noise scheduling coefficient for each diffusion step. From step 1 to step 2 The cumulative noise retention factor of the step; Indicates the first The noisy sequence obtained by the step; Indicates by parameters The noise predicted by the controlled causal-guided denoising network; For causal enhancement representation matrix, This is the optimized situational cause-effect graph; the loss is achieved by minimizing the true noise. With prediction noise The differences between them enable the model to learn how to recover high-quality sequence samples from noisy sequences; The causal structure preservation loss ensures that the generated sequence and the original sequence maintain consistent dependencies on the causal graph; it is defined as follows by comparing the difference in mutual information between the generated sequence and the original sequence on causal edges: (10) in, Preserve loss for causal structure; For the optimized situational cause-effect diagram The set of causal edges in; Representing variables With variables There is a causal dependency between them; and These represent the nth sequence in the generated sequence. The first variable and the second Time series of 1 variable; and They represent the nth and nth digits in the real sequence, respectively. The first variable and the second Time series of 1 variable; The mutual information estimate is used to measure the statistical dependence strength between two variable sequences; Equation (10) constrains the mutual information difference between the generated sequence and the real sequence on the causal edge, so that the generated situation sequence maintains a causal dependence structure similar to the real data. The multimodal alignment loss ensures that the generated sequence remains consistent with the original sequence at the multimodal feature level, and is defined as: (11) in, For multimodal alignment loss; The sequence length; The number of modes; Indicates the first Types of modes, Indicates the actual sequence at time step The Modal characteristics; Indicates the generated sequence at time step The Modal characteristics; Indicates the first The encoder corresponding to each mode; Let L be the square of the L2 norm; Equation (11) constrains the distance between the generated sample and the real sample in each modality coding space, so that the generated sequence maintains semantic consistency at the multi-source information level; The total loss function is the weighted sum of the three: (12) in, The total loss function for model training; The diffusion reconstruction loss is used to constrain the accuracy of noise prediction. The loss is used to preserve the causal structure and constrain the consistency between the generated sequence and the real sequence in terms of causal dependencies. This is a multimodal alignment loss used to constrain the consistency between the generated sequence and the real sequence in different modal feature spaces; and To balance the hyperparameters, the weights of the causal structure preservation loss and the multimodal alignment loss in the total loss are adjusted respectively; After training, the inference process for generating new sequences is as follows: First, initial noise is sampled from a standard Gaussian distribution. This noise can be understood as the first The initial state of the random sequence at each diffusion step; then following the reverse diffusion process from... Iterate step by step to Each step utilizes a causal guided denoising network (CGDT) to predict the noise components in the current noisy sequence, and combines this with a causal enhancement representation matrix. With the optimized situational cause-effect diagram Update the current sequence: (13) in, Indicates the first Noisy sequence states of one reverse diffusion step; Indicates by The noisy sequence state obtained from the previous time step after denoising; This represents the total number of diffusion steps. Number the current backdiffusion step; For the first The noise scheduling coefficient of the step; This is the cumulative noise retention factor; For denoising networks under certain conditions Multimodal fusion features Cause and effect diagram of situation Noise predicted under constraints; For the first The standard deviation of the random noise added in each step; The noise is standard Gaussian random noise; the meaning of this process is: the model first generates the first... The random noise sequence is processed step by step, and then further denoised under the guidance of causal enhancement representation and situational causal graph until a clean sequence at step 0 is obtained; finally This is the generated situation sequence sample.
8. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 7, characterized in that, In the privacy protection and real-time dynamic generation stage, streaming diffusion generation includes: Let the trained multimodal fusion diffusion model be... The goal of this step is to enable the real-time generation of new sequence samples using the model while protecting the privacy of the original training data, and to support online updates and fast inference in dynamic scenarios. Employing a differential privacy training strategy based on causal perception, utilizing situational causal graphs. The variables in the training samples, along with their corresponding model input channels, encoding channels, and related parameter gradients, are grouped and pruned to reduce utility loss due to noise. Based on the correlation strength of nodes in the causal graph, the variables are divided into several causal clusters, and the model parameter gradients processing these variables are assigned to the corresponding causal clusters, with parameters within each cluster sharing a privacy budget. For the ... For a causal cluster, the upper limit of sensitivity after gradient clipping is: (14) in, For the first Upper limit of gradient sensitivity for a causal cluster; and These are adjacent datasets, differing by only one situation sequence sample; All parameters of the model; In order to be with the first A subset of model parameters corresponding to each causal cluster; The model parameters are: The training dataset is The total loss calculated at that time; Represents the total loss relative to a subset of parameters. The gradient; Denotes the 2-norm; Equation (14) indicates that between any two adjacent datasets, the first norm is... The maximum magnitude of gradient change corresponding to each causal cluster; Subsequently, Gaussian noise is added to the clipped gradient of each causal cluster to achieve... - Differential privacy protection: (15) in, For the first The gradient of a causal cluster after adding differential privacy noise; For the first The original gradients of the parameters corresponding to each causal cluster; This is used to clip gradients, ensuring that their L2 norm does not exceed the sensitivity upper bound. ; With a mean of 0 and a covariance of Gaussian noise; For noise scale, determined by privacy budget The size of the causal cluster is determined together with the size of the cluster. It is an identity matrix, and its dimension is the same as the first one. Each causal cluster corresponds to a gradient vector with the same dimension, which is used to represent the addition of mutually independent homoscedastic Gaussian noise in each gradient dimension; through Equation (15), each causal cluster is assigned an independent amount of noise. The more tightly packed the causal structure of the cluster, i.e. the critical dependency path, the smaller the noise is assigned, thereby protecting the generation quality of the critical causal link. The allocation of privacy budgets follows the causal importance criterion: defining causal edges. Importance weight ,in For the variables in formula (1) right Granger regression coefficients, If the lag order is the largest, then the causal cluster Privacy Budget Positively correlated with the average importance of edges within the cluster: (16) in, For the total privacy budget, To be assigned to the Privacy budget for a causal cluster For the first The set of causal edges contained in a causal cluster For the optimized situational cause-effect diagram The set of all causal edges in the array. Causal edge Importance weights; Equation (16) calculates the total privacy budget based on the proportion of edge weights within the causal cluster to all edge weights. Assigned to different causal clusters; through adaptive assignment, key causal paths obtain a more reasonable privacy budget configuration, thereby achieving a balance between privacy protection and generation quality.
9. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 8, characterized in that, In the privacy protection and real-time dynamic generation phase, the state caching mechanism includes: Let the current actual time be... The generation state is That is, at actual time The intermediate representations of each diffusion step in the cache; where, This refers to the total number of diffusion steps or the total number of denoising steps in the diffusion model. Indicates the first One diffusion step; Indicates actual time Next The intermediate sequence states corresponding to each diffusion step; when new data... Upon arrival, only partial updates are needed based on the existing state: (17) in, For the first The update step size coefficient for each diffusion step. For lightweight network updates, Indicates actual time Newly arrived situation increment data, This is the optimized situational cause-effect graph; the lightweight update network is based on... Determine the variables and causal paths affected by the new data, and only update the local diffusion state affected by the new data.
10. The sequence data diffusion generation method for long-dependency structure modeling as described in claim 9, characterized in that, In the privacy protection and real-time dynamic generation phase, a joint optimization objective of privacy-utility-delay is established, and the optimal privacy budget allocation and diffusion step configuration are adaptively selected, including: Define generation utility To generate a similarity between the sample and the original data distribution, delay Privacy costs for average generation time Total privacy budget consumption; optimization objective is to maximize the weighted composite score: (18) in, The evaluation objectives are based on a comprehensive assessment of privacy, utility, and delay. This represents the privacy budget allocation scheme for each causal cluster; This represents the total number of denoising steps in the diffusion model; The generation utility, given a privacy budget allocation and number of diffusion steps, is measured by the similarity between the generated samples and the distribution of real-world situational data. This represents the generation delay, which typically varies with the number of diffusion steps. Increase as it grows; This indicates penalties related to the strength of privacy protection, when the privacy budget... When the value is smaller, privacy protection is stronger, but model training noise is greater. This term is used to characterize the impact of privacy protection on generative utility. and These are the delay penalty coefficient and the privacy penalty coefficient, respectively; Equation (18) means that, under the premise of meeting the privacy protection requirements of sensitive data, the privacy budget of each causal cluster is jointly selected. and diffusion steps The goal is to maximize generation utility while controlling generation latency and quality loss due to privacy protection. This optimization problem is solved using Bayesian optimization, which automatically selects the optimal privacy budget allocation strategy and diffusion steps based on the specific application scenario.