A Multi-Perspective Parallel Diffusion Time Series Generation Method

Abstract

This invention discloses a multi-perspective parallel diffusion time series generation method, comprising: performing diffusion modeling on the original multivariate time series to construct noisy input samples; extracting the basic time series representation through a time-perspective backbone encoder; extracting variable relationship perspective features through a fluctuation-aware graph module, constructing graph structure relationships between variables and performing message passing; extracting frequency domain perspective features through a dual-mode frequency domain module, adaptively selecting a structure-preserving mode or a learnable transformation mode for frequency domain processing based on spectral complexity; and performing parallel fusion of the three perspective features (time perspective, variable relationship perspective, and frequency domain perspective) and completing sequence generation through a decoder. This invention, through a multi-perspective parallel denoising coding mechanism, simultaneously utilizes three complementary types of information during the diffusion generation process: long-term time dependence, inter-variable correlation, and global periodicity, effectively improving the generation quality of multivariate long-term time series and possessing good innovation, practicality, and promotional value.

Description

Technical Field

[0001] This invention belongs to the field of time series generation and artificial intelligence technology, and in particular relates to a multi-view parallel diffusion time series generation method. Background Technology

[0002] Time series generation has significant application value in fields such as financial risk control, industrial manufacturing, clinical medicine, traffic scheduling, and power and energy. Especially in scenarios with insufficient training samples, privacy restrictions, or anomalous samples, generating high-quality, multivariate, long-term time series data is crucial for improving the generalization capabilities of downstream anomaly detection, missing value imputation, prediction, and decision-making models.

[0003] Existing time series generation technologies have generally gone through the following stages.

[0004] (1) Traditional statistical methods stage: Early methods mainly included statistical models such as ARIMA and Kalman filtering. These methods have certain effects on low-dimensional, regular, and stationary series, but they are difficult to capture deep time series dependencies and complex variable interactions for high-dimensional, nonlinear, non-stationary, and multivariate long-term time series that are common in reality, so their generation ability is limited.

[0005] (2) Single-step generation stage based on GAN / VAE: Generative adversarial networks and variational autoencoders have driven the development of time series generation. For example, models such as C-RNN-GAN, TimeGAN, DoppelGANger, TTS-GAN, and TimeVAE can model time dependence and feature relationships to a certain extent. However, GAN-type methods generally suffer from problems such as training instability and mode collapse, while VAE-type methods are prone to phenomena such as sample blurring and missing details. These problems are more obvious in multivariate long sequence generation tasks.

[0006] (3) Iterative generation stage based on diffusion model: In recent years, diffusion model has gradually become an important direction for time series generation due to its strong distribution fitting ability and relatively stable training process. Although diffusion model can approximate complex data distribution through progressive noise addition and reverse denoising, for multivariate long-term time series, as the number of denoising iterations increases, the model often finds it difficult to simultaneously maintain long-term time dependence, inter-variable correlation and global periodicity. It is prone to problems such as overall trend shift, weakening of variable relationship and disordered periodic rhythm, which leads to degradation of the quality of generated data.

[0007] Existing technologies still have the following common shortcomings in the task of generating multivariate long-term time series:

[0008] (1) Insufficient ability to model dynamic changes in variable relationships.

[0009] While existing graph modeling methods can learn the dependencies between variables to some extent, they lack an effective description of the fluctuating characteristics of variable relationships over time, making it difficult to stably maintain cross-variable dependencies in non-stationary scenarios.

[0010] (2) The frequency domain modeling method is relatively simple.

[0011] Existing frequency domain methods typically employ a uniform frequency domain modeling strategy, which makes it difficult to adapt to both purely periodic dominant sequences and complex aliased spectral sequences simultaneously, and their adaptability to different types of data is insufficient.

[0012] (3) Insufficient multi-perspective collaborative modeling.

[0013] Existing methods mostly model from a single perspective or use a serial structure to process time, variable relationships and frequency domain information separately, making it difficult to fuse multiple complementary information in parallel during the denoising process, thus limiting further improvement in generation quality.

[0014] Therefore, it is necessary to propose a new method for generating multivariate long-term time series to address the problem that existing diffusion models struggle to simultaneously consider long-term time dependence, inter-variable correlations, and global periodic patterns during the long-term generation process. Summary of the Invention

[0015] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art and proposes a multi-perspective parallel diffusion time series generation method. It aims to solve the problem that existing multivariate long-term time series generation methods are difficult to simultaneously take into account long-term time dependence, inter-variable correlation and global periodicity in the diffusion denoising process, thereby improving the performance of the generated series in terms of statistical distribution consistency, cross-variable correlation and preservation of time series regularity.

[0016] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0017] In a first aspect, the present invention provides a multi-view parallel diffusion time series generation method, comprising the following steps:

[0018] S1: Perform diffusion modeling on the original multivariate time series to construct noisy input samples;

[0019] S2: Input the noisy input samples obtained in S1 into the time-view backbone encoder to model the long-term dependencies across time steps and obtain the basic time series representation;

[0020] S3: Input the basic time series features obtained in S2 into the fluctuation perception graph module to construct the graph structure relationship between variables and output the variable relationship perspective features. The fluctuation perception graph module describes the dependency strength between variables through the adjacency matrix based on the basic time series representation and performs message passing on the graph structure.

[0021] S4: Input the basic time-series features obtained in S2 into the dual-mode frequency domain module to model the global periodicity in the sequence and output the frequency domain perspective features; The dual-mode frequency domain module performs a Fourier transform on the input features along the time dimension, maps the long-range periodic dependence in the time domain to the frequency domain representation, and adaptively selects the structure-preserving mode or the learnable transform mode for frequency domain processing according to the spectral complexity.

[0022] S5: The results of S2, S3 and S4 are fused in parallel to obtain the fused multi-view encoded representation. The fused multi-view encoded representation is then input to the decoding end to obtain the generated multivariate long-term time series sample.

[0023] Furthermore, S1 includes:

[0024] The original multivariate time series samples are obtained. During the forward diffusion stage, Gaussian noise is gradually added according to a preset noise scheduling strategy to obtain the noisy sample at the current diffusion time step. During the reverse diffusion stage, the noisy sample and its corresponding diffusion time step are used as the input of the denoising network to output the estimated value of the original sample. The state of the previous time step is then restored based on the estimated value.

[0025] Furthermore, the wave sensing graph module in step S3 includes the following sub-steps:

[0026] Linear projection is performed on the basic time series features to align the feature dimensions with the variable dimensions;

[0027] Calculate the similarity between variables for the dimensionally aligned variable node representations, and obtain the dynamic similarity adjacency matrix through normalization operations;

[0028] Set a normalized prior adjacency matrix as the initial structural constraint for the variable relations;

[0029] Construct a fluctuation sensing and control mechanism: Calculate the mean and standard deviation of the input sequence along the time dimension, and combine the two into a global state vector. Map this global state vector through a node encoder to obtain the sample-level dynamic intensity coefficient. Combine the learnable basic fusion parameters to perform adaptive weighted fusion of the dynamic similarity graph and the prior graph to obtain the final adjacency matrix.

[0030] Perform graph convolution and nonlinear transformation on the final adjacency matrix to output variable relationship perspective features.

[0031] Furthermore, the dual-mode frequency domain module in step S4 includes the following sub-steps:

[0032] Perform a real-valued Fast Fourier Transform along the time dimension on the input features to map the time-domain signal to a frequency-domain representation;

[0033] Calculate the spectral entropy as a measure of spectral complexity.

[0034] When the spectral entropy is lower than a preset threshold, the structure preservation mode is adopted: frequency-channel selection weights are generated through a gating network to selectively recalibrate the spectral amplitude.

[0035] When the spectral entropy is higher than or equal to a preset threshold, a learnable transformation mode is adopted: the real and imaginary parts of the spectrum are respectively learned and mapped, and then recombined into a new complex frequency domain representation.

[0036] The frequency domain representation is mapped back to the time domain through inverse Fourier transform, and then the frequency domain viewpoint features are output after normalization and learnable gated injection.

[0037] Furthermore, the multi-view parallel fusion in step S5 adopts a gated residual mechanism, which adaptively adjusts the contribution ratio of each view information through learnable gate parameters to obtain the fused multi-view encoded representation.

[0038] Furthermore, the decoding end uses self-attention and cross-attention joint modeling, and combines trend term and seasonal term decomposition to refine the sequence representation of the current diffusion step, outputs the original sequence estimate, further derives the state of the previous time step based on the estimate, and repeats steps S2 to S5 until all back diffusion iterations are completed, finally obtaining the generated multivariate long-term time series sample.

[0039] Furthermore, the time-view backbone encoder in step S2 adopts a Transformer encoding structure, with noisy sequences and time step embedding information as inputs and basic temporal features as outputs.

[0040] Furthermore, in the aforementioned fluctuation sensing and control mechanism, the global state vector simultaneously encodes the overall level and fluctuation intensity of the current sample; the dynamic intensity coefficient reflects the credibility of the current dynamic graph.

[0041] In a second aspect, the present invention provides an electronic device, including a processor and a memory communicatively connected to the processor and used to store executable instructions of the processor, wherein the processor is used to execute the above-described multi-view parallel diffusion time series generation method.

[0042] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned multi-view parallel diffusion time series generation method.

[0043] Compared with existing technologies, the multi-view parallel diffusion time series generation method described in this invention has the following advantages:

[0044] This invention achieves collaborative modeling of long-term time dependencies, inter-variable correlations, and global periodic patterns within a diffusion generation framework by constructing a time-perspective backbone encoder, a fluctuation sensing map module, a dual-mode frequency domain module, and a multi-perspective parallel fusion mechanism. Compared to existing single-perspective or serial generation methods, the method described in this invention can more effectively improve the generation quality of multivariate long-term time series, demonstrating significant innovation and practical value. Attached Figure Description

[0045] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0046] Figure 1 This is a schematic diagram of the overall framework of the present invention;

[0047] Figure 2 This is a schematic diagram of the wave sensing graph module of the present invention;

[0048] Figure 3 This is a schematic diagram of the dual-mode frequency domain module of the present invention;

[0049] Figure 4 This is a schematic diagram of the overall process of the present invention;

[0050] Figure 5 This is a visualization comparison of the generated samples and real samples using methods such as PCA in this invention.

[0051] Figure 6 This is a visualization comparison of the generated samples and real samples using the t-SNE method in this invention;

[0052] Figure 7 This is a visualization comparison of the generated samples and real samples using the kernel density estimation method of this invention;

[0053] Figure 8 This is a thermal diagram of the spectral structure of the present invention. Detailed Implementation

[0054] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0055] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] This invention proposes a multi-perspective parallel diffusion time series generation method (TriDiff) that integrates wave perception maps and dual-mode frequency domain. It takes the reverse stepwise denoising process of the diffusion model as the core, constructs a three-way parallel feature extraction structure with time perspective, variable relationship perspective and frequency domain perspective at the encoding end, and then completes multi-perspective joint modeling through an adaptive fusion mechanism. Finally, the generated result is output at the decoding end.

[0057] In the task of generating multivariate long-term time series data, the training dataset consists of multiple samples, each of length [missing information]. The variable dimension is A multivariate time series segment. For any sample, its original sequence can be represented as... The goal of this invention is to learn a diffusion generation model that can map Gaussian noise into a synthesized sequence through a gradual backdiffusion process, and make the distribution of the generated sequence gradually approximate the distribution of the real data. During training, a noisy sample is input at the current diffusion step, and the estimated value of the original sample is directly predicted. The state of the previous time step is then recovered from this estimate. This process is repeated iteratively until the backdiffusion is completed.

[0058] like Figure 1 As shown, the overall model of this invention is constructed based on the Transformer diffusion framework of a decoupled time representation encoder-decoder. The denoising network receives the noisy sequence and time step information as input at each discrete time step t, and outputs an estimate of the original sequence. Specifically, the encoder introduces a fluctuation-aware graph module and a dual-mode frequency domain module on top of the time modeling backbone, forming a multi-view parallel denoising encoder. Subsequently, the three features are adaptively fused through a gated residual mechanism. The decoder adopts a DiffusionTS-like structure, refining the sequence representation through self-attention, cross-attention, and trend and seasonal term decomposition, thereby completing the generation of multivariate long-term time series.

[0059] The specific implementation steps of the present invention are described in detail below, such as... Figure 4 As shown:

[0060] S1. Perform diffusion modeling on the original multivariate time series to construct noisy input samples.

[0061] First, obtain the original multivariate time series samples and represent them as a series of length [length missing]. The variable dimension is The sequence segment. During the forward diffusion phase, Gaussian noise is gradually added to the original sequence according to a preset noise scheduling strategy to obtain the current diffusion time step. Noisy samples During the back diffusion phase, the noisy sample... and its corresponding diffusion time step As input to the denoising network, the output is a response to the original sample. The estimated value Then, based on this estimate, the state of the previous time step can be further restored. This allows for gradual noise reduction and generation.

[0062] In practical implementation, noisy input samples can be represented as: ,in Indicates batch size. Indicates the length of time. Indicates the dimension of the variable. Time step By inputting the data into the denoising network via temporal embedding, the model becomes aware of the current denoising stage. This step provides a unified input for subsequent multi-view feature extraction at the encoding end.

[0063] S2. Extracting the basic temporal representation through a time-perspective backbone encoder.

[0064] The noisy input samples obtained in step S1 are input into the time-view backbone encoder to model the long-term dependencies across time steps and obtain the basic temporal representation. The backbone encoder employs a Transformer coding structure to extract sequence representations from a global temporal perspective, capturing the main temporal dynamic features of the sequence at the current diffusion step.

[0065] In this invention, the input to the time-view backbone encoder is a noisy sequence. Including time step embedding information, the output is the basic time series feature. ,in To hide the representation dimension, this basic temporal feature is not the final output, but rather serves as the common input to the subsequent fluctuation sensing map module and the dual-mode frequency domain module, enabling three-way parallel feature extraction.

[0066] Unlike methods that rely solely on single-path time modeling, this invention uses a temporal backbone as a common foundation for multi-view modeling: on the one hand, the backbone encoder is responsible for extracting long-term temporal dependencies; on the other hand, graph branches and frequency domain branches enhance the basic representation from two supplementary perspectives—variable correlation and global periodic structure—thereby providing more complete information support for subsequent inverse denoising.

[0067] S3. Extract variable relationship perspective features through the fluctuation perception graph module.

[0068] The basic time series features obtained in step S2 Input the fluctuation perception graph module to construct the graph structure relationship between variables and output the variable relationship perspective features. The fluctuation-aware graph module treats each variable as a node in the graph, explicitly describes the dependency strength between variables through an adjacency matrix, and performs message passing on the graph structure to achieve cross-variable information aggregation and structural enhancement.

[0069] Specifically, such as Figure 2 As shown, the wave sensing graph module includes the following sub-processes:

[0070] First, dimensional alignment is performed on the basic temporal features. This is because the hidden representation has different dimensions. Each variable needs to be assigned an independent subspace. This invention first performs a linear projection on the input features so that the feature dimension can be aligned with the variable dimension, thereby so that each variable corresponds to an independent sub-representation space.

[0071] Secondly, a dynamic similarity graph is constructed. The similarity between variables is calculated for the dimensionally aligned variable nodes, and a dynamic similarity adjacency matrix is ​​obtained through normalization, which represents the dynamic dependencies between variables in the current sample.

[0072] Third, a prior graph is introduced. To avoid instability of the dynamic graph in the early stages of training, this invention sets a normalized prior adjacency matrix as an initial structural constraint for variable relationships. This prior graph, together with the dynamic similarity graph, participates in subsequent fusion.

[0073] Fourth, a fluctuation-aware control mechanism is constructed. The mean and standard deviation of the input sequence are calculated along the time dimension and combined into a global state vector. This global state vector simultaneously encodes the overall level and fluctuation intensity of the current sample, characterizing the degree of dynamic change of the sequence at the current diffusion step. Subsequently, the global state vector is mapped through a node encoder to obtain sample-level dynamic intensity coefficients, reflecting the credibility of the current dynamic graph. Then, combined with learnable basic fusion parameters, an adaptive weighted fusion of the dynamic similarity graph and the prior graph is performed to obtain the final adjacency matrix.

[0074] Fifth, graph convolution and nonlinear transformation are performed on the final adjacency matrix. For each sample and each time step, the adjacency matrix is ​​used for message passing in the variable dimension to aggregate information from other variable nodes. Subsequently, the expressive power is further enhanced through shared linear layers and activation functions, and the gated residual form is fused with the original representation to output variable relationship-based features. .

[0075] In this invention, the input to the wave sensing graph module is... The output is Through the above structure, the model can adaptively balance prior structural information and dynamic correlation information according to the degree of sample fluctuation, thereby enhancing the ability to maintain the stability of cross-variable dependencies during the generation of multivariate long-term time series.

[0076] S4. Extract frequency domain view features using a dual-mode frequency domain module.

[0077] The basic time series features obtained in step S2 The dual-mode frequency domain module is input to model the global periodicity in the sequence and outputs frequency domain features. Dual-mode frequency domain module, such as Figure 3 As shown, a real-valued fast Fourier transform is first performed on the input features along the time dimension to map the long-range periodic dependencies in the time domain to the frequency domain representation, so as to capture global periodic patterns at a lower cost.

[0078] Considering the significant differences in the spectral structure of different data, this invention sets two frequency domain modeling modes based on spectral complexity and adaptively selects between them:

[0079] The first type is the structure-preserving mode. For purely frequency-dominant or phase-sensitive data, only the spectral amplitude is selectively recalibrated without explicitly changing the phase structure, thereby preserving the original periodic waveform as much as possible. This mode generates frequency-channel selection weights through a gating network and adjusts the spectral amplitude, making it suitable for data with a clear periodic structure and concentrated dominant frequency.

[0080] The second approach is the learnable transformation mode. For data with multiple peaks in the spectrum, strong noise, or more complex spectral structures, learnable mappings are performed on the real and imaginary parts of the spectrum to extract more generalized frequency combination information. After the real and imaginary parts are transformed, they are then recombined into a new complex frequency domain representation.

[0081] Regarding the selection of the two modes, this invention introduces spectral entropy as a metric for spectral complexity: when the spectral entropy is low, it indicates that the spectral energy is mainly concentrated in a few major frequency regions, and the structure-preserving mode is preferred; when the spectral entropy is high, it indicates that the spectral distribution is more complex, and the learnable transformation mode is preferred. This setting enables the frequency domain modeling method to adaptively adapt to different data scenarios.

[0082] After completing the frequency domain processing, the frequency domain representation is mapped back to the time domain using an inverse Fourier transform to obtain the frequency domain branch output. To avoid excessive perturbation of the backbone representation by the frequency domain branch in the early stages of training, this invention further employs a normalization and learnable gating injection strategy to normalize the frequency domain output and control the intensity of frequency domain information introduction through learnable injection coefficients, ultimately outputting frequency domain perspective features. .

[0083] In this invention, the input of the dual-mode frequency domain module is The output is By employing the aforementioned dual-mode frequency domain modeling scheme, the model can both maintain its periodic structure and adapt to complex spectral data, thereby improving the ability to model global periodic patterns in the generation process of multivariate long-term time series.

[0084] S5. Perform parallel fusion of multi-view features and generate the sequence via a decoder.

[0085] The temporal backbone features obtained in step S2, the variable relationship features obtained in step S3, and the frequency domain features obtained in step S4 are fused in parallel. Specifically, using the temporal features as the basic representation, the variable relationship features and the frequency domain features are injected into the backbone representation using a gated residual method. The contribution ratio of each perspective information is adaptively adjusted by a learnable gating parameter to obtain the fused multi-view encoded representation.

[0086] The encoded representation obtained through multi-view fusion serves as the input to the decoder. The decoder employs an encoder-decoder Transformer structure, using joint modeling with self-attention and cross-attention, and incorporating trend and seasonal term decomposition to refine the sequence representation of the current diffusion step, outputting the original sequence estimate. Then, based on the estimated value, the state of the previous time step is further derived. Then, repeat steps S2 to S5 until all back diffusion iterations are completed, and finally obtain the generated multivariate long-term time series samples.

[0087] In this invention, the input to the multi-view fusion module includes , and The output is the fused encoded representation; the input to the decoding generation module is the fused encoded representation and the current diffusion step information, and the output is the original sample estimate. And the generated sequence is gradually recovered. Through the above-mentioned parallel fusion and stepwise decoding mechanism, this invention can simultaneously utilize three complementary types of information—time dependence, variable correlation, and frequency domain regularity—during the diffusion denoising process, thereby improving the authenticity, correlation, and temporal consistency of the generated sequence.

[0088] To verify the effectiveness of the TriDiff proposed in this invention in multivariate long-term generation tasks, this invention conducted a large number of benchmark experiments, systematically evaluated the prediction performance of the method using multiple public datasets, and compared it with existing mainstream generation models.

[0089] (1) Experimental environment and setup

[0090] The method of this invention is implemented in a deep learning environment based on the PyTorch framework. Both training and inference are performed on a device equipped with an NVIDIA GeForce RTX 2080Ti GPU (11GB VRAM). Five public datasets were selected for the experiments: Stocks, Energy, fMRI, MuJoCo, and Sines. See Table 1 for details.

[0091] Table 1

[0092]

[0093] To evaluate the performance of the method of this invention, DiffusionTS, TimeVAE, TimeGAN, and DiffWave were selected as baseline models for comparison. All models were implemented in the same hardware environment and under a unified training framework to ensure the comparability of experimental results. In the long-term time series generation experiment, this invention set two typical window lengths of 128 and 256 to examine the performance changes of the model in medium-length and longer-length generation tasks.

[0094] The main hyperparameters of the method in this invention are set according to the data size and complexity on different datasets. For example, the number of attention heads is uniformly set to 4; the hidden dimension head_dim is set to 16 or 24 depending on the dataset; the number of encoder layers is adjusted between 1 and 4, and the number of decoder layers is adjusted between 2 and 4; the batch size is set to 64 or 128 depending on the data complexity; the number of diffusion steps is set between 500 and 1000; and the number of training steps is set to 10000, 12000, 14000, 15000, or 25000 depending on the dataset size. Through the above settings, the model can maintain good training stability and generation effect on different datasets.

[0095] To comprehensively evaluate the quality of generated sequences from different perspectives, this invention uses four indicators as evaluation criteria: Context-FID score, Correlational score, Discriminative score, and Predictive score. The lower the value of these indicators, the closer the generated sample is to the real sample, and the better the model's generation performance.

[0096] The contextual FID score measures the similarity in statistical distribution between the generated sequence and the real sequence. This metric replaces the Inception feature extraction network in the traditional image domain FID with the time series representation model TS2Vec. After encoding the real and generated samples into a unified contextual representation space, the distribution difference between the two is calculated, thus better characterizing the overall distribution approximation of the time series generation results.

[0097] The correlation score measures the consistency between generated and real data in terms of dependencies between variables. This metric assesses the model's ability to preserve cross-variable correlation structures by calculating the difference in the covariance matrices between pairs of variables in the real and generated data. For multivariate long-term time series generation tasks, this metric directly reflects whether the model effectively reproduces the interactions between variables.

[0098] The discrimination score measures the degree to which real samples and generated samples can be distinguished. Specifically, by training a discriminator to differentiate between real and generated samples, if the model's generated results are closer to the true distribution, the discriminator's accuracy is closer to random guessing, and the corresponding discrimination score is lower. Therefore, this metric reflects the authenticity of the generated samples.

[0099] The prediction score is used to examine whether the generated data exhibits transferable consistency with temporal patterns. Based on the TSTR evaluation process, this metric calculates the mean absolute error of the prediction model trained using generated samples on the real test set, thereby assessing the ability of the generated data to support downstream prediction tasks. A lower metric indicates that the generated data better reproduces predictable patterns in the real data.

[0100] (2) Experimental results and comparison results

[0101] The proposed TriDiff method was systematically compared with DiffusionTS, TimeVAE, TimeGAN, and DiffWave on five public datasets and two generation length settings. Experimental results show that the proposed TriDiff method outperforms existing comparative models on most key metrics, indicating that it can better capture long-term time dependencies, inter-variable correlations, and global periodic patterns simultaneously, thereby effectively improving the generation quality of multivariate long-term time series data.

[0102] Overall, compared to the suboptimal results, the method of this invention reduces the performance of Context-FID, Correlational, and Discriminative metrics by 23.63%, 18.84%, and 10.90%, respectively, with a total reduction of 14.08%, as shown in Table 2. These results demonstrate that the multi-view parallel denoising encoder constructed in this invention can improve the generation effect across multiple dimensions, including statistical distribution, correlation structure, and sample realism.

[0103] Table 2

[0104]

[0105] Specifically, on the Energy dataset, the method of this invention achieves Context-FID of 0.0614±0.007, Correlational of 0.4462±0.103, Discriminative of 0.1636±0.077, and Predictive of 0.2472±0.000 when generating a length of 128; and 0.0853±0.009, 0.3333±0.050, 0.2103±0.104, and 0.2459±0.000 respectively when generating a length of 256, which is better than most of the comparison models. This indicates that the method of this invention has good generation capabilities on high-dimensional, strongly correlated, and noisy periodic data.

[0106] On the MuJoCo dataset, the method of this invention achieves Context-FID values ​​of 0.0702±0.006, Correlational values ​​of 0.1878±0.027, Discriminative values ​​of 0.0615±0.017, and Predictive values ​​of 0.0058±0.000 when the generation length is 128; and 0.1439±0.015, 0.2185±0.008, 0.0555±0.008, and 0.0053±0.001 respectively when the generation length is 256, demonstrating a strong ability to maintain temporal regularity in the dynamic evolution scenario of multi-physical systems.

[0107] The proposed method also performs well on the Sines and Stocks datasets. Particularly on the Sines dataset, due to its distinct periodicity, the proposed method, through its dual-mode frequency domain mechanism, better preserves the periodic structure, resulting in superior performance in metrics such as Context-FID, Correlational, and Discriminative. For the Stocks dataset, the proposed method also outperforms most comparative models in terms of contextual distribution consistency and variable correlation structure reproduction, demonstrating its applicability to complex financial time series data.

[0108] In addition to quantitative indicators, this invention also employs methods such as PCA, t-SNE, and kernel density estimation (KDE) to visually compare the generated samples with the real samples, for example... Figure 5-7 As shown, taking the generated Sines dataset of length 128 as an example, the method of this invention has a more sufficient overlap with the distribution of real samples in the low-dimensional embedding space, and the position and peak details of the KDE main peak are also closer to the real distribution, further illustrating that the data generated by the method of this invention is more consistent with the real data in terms of feature structure and statistical distribution.

[0109] (3) Module effectiveness analysis (ablation experiment)

[0110] To further verify the effectiveness of the wave-aware graph mechanism and the dual-mode frequency domain mechanism in this invention, systematic ablation experiments were conducted on the Energy and Sines datasets. The Energy dataset served as the primary ablation validation dataset, used to analyze the contributions of the wave-aware graph module and the dual-mode frequency domain module to the overall generation quality. The Sines dataset, as a period-dominated dataset, was used to focus on analyzing the role of the dual-mode frequency domain mechanism under different spectral complexity scenarios.

[0111] For the Energy dataset, this invention sets up four sets of comparative models: FAGraph+MLP (replacing the graph module with a simple linear module), w / o FA (removing the wave-aware component), DMFreq+MLP (replacing the dual-mode frequency domain module with a simple linear module), and the complete TriDiff model. Overall, the complete TriDiff model outperforms the other reduction schemes in Context-FID, Correlational, and Discriminative metrics, verifying the effectiveness of the wave-aware graph module and the dual-mode frequency domain module in improving generation quality. (See Table 3.)

[0112] Table 3

[0113]

[0114] Further analysis reveals that removing the fluctuation-aware module degrades the model's performance on the Energy dataset across multiple metrics, indicating that relying solely on static prior maps or simple dynamic maps is insufficient to adequately adapt to the dynamic changes in variable relationships within non-stationary scenarios. Furthermore, replacing DMFreq with a simple linear module weakens the model's ability to model frequency domain patterns, leading to a decrease in generation performance. Although the DMFreq+MLP performs slightly better than the complete TriDiff model on the Predictive Score for a generation length of 256, the difference is small and does not affect the overall advantage of the complete model. This phenomenon suggests that the fluctuation-aware map module still has room for further optimization under longer sequence generation conditions, but its overall technical performance remains positive.

[0115] For the Sines dataset, this invention sets up three ablation schemes: with / o Freq (removing the frequency domain mechanism), with / o SPM (removing the structure-preserving mode), and with / o LFT (removing the learnable transform mode), and compares them with the complete model. Experimental results show that removing the dual-mode frequency domain mechanism significantly degrades model performance. For example, with / o Freq, the Context-FID reaches 0.7418±0.098 with a generation length of 256, which is significantly worse than the complete TriDiff. Furthermore, removing the structure-preserving mode and the learnable transform mode respectively also results in varying degrees of degradation in Correlational, Discriminative, and Predictive metrics. This indicates that the two modes in the dual-mode frequency domain mechanism function differently for different spectral morphologies, jointly improving the generation quality in periodic signal scenarios.

[0116] Furthermore, to further verify the differences in frequency domain characteristics between different datasets, this invention plotted spectral heatmaps for the Sines and Energy datasets, as shown below. Figure 8 As shown.

[0117] The results show that the energy distribution in the Energy dataset is spread across multiple frequency positions, exhibiting a complex spectral structure and high spectral entropy; while the spectrum of the Sines dataset is mainly concentrated in a few low-frequency regions, with a clear periodic structure and low spectral entropy. These results experimentally demonstrate the differentiated requirements for frequency domain modeling methods across different datasets, and also indirectly verify the rationality and necessity of the dual-mode frequency domain mechanism proposed in this invention.

[0118] In summary, this invention effectively improves the overall performance of multivariate long-term time series generation tasks by introducing a fluctuation-aware graph mechanism and a dual-mode frequency domain mechanism at the encoding end of the diffusion model, and by employing a multi-view parallel fusion approach. Comparative experiments verify the effectiveness of the proposed method on multiple public datasets, and ablation experiments further demonstrate that the fluctuation-aware graph module, the dual-mode frequency domain module, and their internal components all play important roles in improving generation quality. Therefore, this invention possesses significant technical advancement and practical application value.

[0119] The proposed TriDiff method, a multi-perspective parallel diffusion time series generation method incorporating fluctuation sensing maps and dual-mode frequency domain, addresses the shortcomings of existing multivariate long-term time series generation models in terms of preserving long-term time dependencies, modeling inter-variable correlations, and restoring global periodic patterns. By introducing fluctuation sensing map mechanisms, dual-mode frequency domain mechanisms, and multi-perspective parallel fusion structures into the diffusion denoising network, it effectively improves the overall quality of generated sequences in terms of statistical distribution consistency, structural consistency, and temporal pattern consistency, demonstrating significant technical effects and practical application value.

[0120] (1) It has the ability to collaboratively model from multiple perspectives, which can simultaneously enhance the joint expression of time dependence, variable correlation and frequency domain regularity.

[0121] This invention constructs a three-way parallel structure at the encoding end of the diffusion model: a time-view backbone, a fluctuation-aware graph branch, and a dual-mode frequency domain branch, moving beyond the limitations of single-view or serial modeling methods in existing approaches. Through this multi-view parallel mechanism, the model can simultaneously utilize long-term dependency information across time steps, cross-variable correlation information, and global periodicity information within the same denoising step. This makes the generated samples more closely resemble real data in terms of overall trend, variable correlation structure, and periodic patterns, thereby improving the overall quality of multivariate long-term time series generation.

[0122] (2) It has the ability to model variable relationships with fluctuation perception, which can improve the stability of cross-variable dependency maintenance in non-stationary scenarios.

[0123] This invention introduces a fluctuation-aware graph module, treating each variable as a node in the graph, and combines a dynamic similarity graph with a priori graph to model variable relationships. Furthermore, instead of simply using a fixed graph structure or a typical dynamic graph structure, this invention constructs a global state vector using the mean and standard deviation of the sequence over time, and generates sample-level dynamic intensity coefficients based on this vector, adaptively adjusting the fusion ratio of prior structural information and dynamic correlation information. Through this design, the model can better adapt to the fluctuations in variable relationships with state changes in non-stationary long sequences, mitigating the degradation in generation quality caused by unstable variable relationships and enhancing the ability of the generated sequence to reproduce the correlation structure of real variables.

[0124] (3) It has the ability to adapt frequency domain modeling for different spectral morphologies, and can take into account both periodic structure preservation and complex spectrum enhancement.

[0125] This invention proposes a dual-mode frequency domain module that performs a Fourier transform on the input representation along the time dimension and adaptively selects either a structure-preserving mode or a learnable transform mode based on spectral complexity. Specifically, for data with a clear dominant frequency and phase sensitivity, the structure-preserving mode is used, selectively recalibrating only the spectral amplitude to preserve the original periodic waveform as much as possible. For data with multi-peak aliasing, strong noise, or more complex structures, the learnable transform mode is used, performing learnable mappings on the real and imaginary parts of the spectrum to extract richer global frequency combination information. This mechanism overcomes the limitations of existing unified frequency domain modeling methods, enabling the model to adopt more appropriate frequency domain enhancement methods for the spectral characteristics of different datasets, thereby improving the ability to model global periodic patterns.

[0126] (4) It has a stronger long-term generation capability, which can alleviate the degradation problem caused by the gradual accumulation of errors in the long sequence generation process of the diffusion model.

[0127] Existing diffusion models often exhibit trend shifts, weakened variable relationships, or disrupted periodic rhythms in multivariate long-term time series generation tasks as the number of backward iterations increases. This invention introduces both variable relationship enhancement and frequency domain regularity enhancement mechanisms into the denoising network, enabling each denoising step to acquire richer and more stable structural information constraints. This allows for more effective maintenance of long-term dependencies, variable coupling relationships, and global periodic regularities during backward generation, mitigating the quality degradation caused by error accumulation and amplification in long-sequence scenarios.

[0128] Example 2:

[0129] An electronic device includes a processor and a memory communicatively connected to the processor and used to store processor-executable instructions, the processor being used to execute the aforementioned multi-view parallel diffusion time series generation method.

[0130] Example 3:

[0131] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned multi-view parallel diffusion time series generation method.

[0132] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this disclosure all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0133] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy and procedures.

[0134] This disclosure is intended to provide implementation schemes for users to selectively prevent the use or access to their personal information data. Specifically, this disclosure is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information is de-identified to protect user privacy.

[0135] The acquisition, transmission, storage, use, and processing of data in this disclosed technical solution all comply with the relevant provisions of national laws and regulations.

[0136] It should be noted that in the embodiments disclosed herein, certain software, components, models, and other existing solutions in the industry may be mentioned. These should be considered as exemplary and are intended only to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used such solutions.

[0137] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0138] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0139] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this disclosure includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of this disclosure pertain.

[0140] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0141] It should be understood that various parts of this disclosure can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0142] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0143] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0144] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present disclosure have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present disclosure. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present disclosure.

Claims

1. A multi-view parallel diffusion time series generation method, characterized in that: Includes the following steps: S1: Perform diffusion modeling on the original multivariate time series to construct noisy input samples; S2: Input the noisy input samples obtained in S1 into the time-view backbone encoder to model the long-term dependencies across time steps and obtain the basic time series representation; S3: Input the basic time series features obtained in S2 into the fluctuation perception graph module to construct the graph structure relationship between variables and output the variable relationship perspective features. The fluctuation perception graph module describes the dependency strength between variables through the adjacency matrix based on the basic time series representation and performs message passing on the graph structure. S4: Input the basic time-series features obtained in S2 into the dual-mode frequency domain module to model the global periodicity in the sequence and output the frequency domain perspective features; The dual-mode frequency domain module performs a Fourier transform on the input features along the time dimension, maps the long-range periodic dependence in the time domain to the frequency domain representation, and adaptively selects the structure-preserving mode or the learnable transform mode for frequency domain processing according to the spectral complexity. S5: The results of S2, S3 and S4 are fused in parallel to obtain the fused multi-view encoded representation. The fused multi-view encoded representation is then input to the decoding end to obtain the generated multivariate long-term time series sample.

2. The multi-view parallel diffusion time series generation method according to claim 1, characterized in that: S1 includes: The original multivariate time series samples are obtained. During the forward diffusion stage, Gaussian noise is gradually added according to a preset noise scheduling strategy to obtain the noisy sample at the current diffusion time step. During the reverse diffusion stage, the noisy sample and its corresponding diffusion time step are used as the input of the denoising network to output the estimated value of the original sample. The state of the previous time step is then restored based on the estimated value.

3. The multi-view parallel diffusion time series generation method according to claim 1, characterized in that: The wave sensing graph module in step S3 includes the following sub-steps: Linear projection is performed on the basic time series features to align the feature dimensions with the variable dimensions; Calculate the similarity between variables for the dimensionally aligned variable node representations, and obtain the dynamic similarity adjacency matrix through normalization operations; Set a normalized prior adjacency matrix as the initial structural constraint for the variable relations; Construct a fluctuation sensing and control mechanism: Calculate the mean and standard deviation of the input sequence along the time dimension, and combine the two into a global state vector. Map this global state vector through a node encoder to obtain the sample-level dynamic intensity coefficient. Combine the learnable basic fusion parameters to perform adaptive weighted fusion of the dynamic similarity graph and the prior graph to obtain the final adjacency matrix. Perform graph convolution and nonlinear transformation on the final adjacency matrix to output variable relationship perspective features.

4. The multi-view parallel diffusion time series generation method according to claim 1, characterized in that: The dual-mode frequency domain module in step S4 includes the following sub-steps: Perform a real-valued Fast Fourier Transform along the time dimension on the input features to map the time-domain signal to a frequency-domain representation; Calculate the spectral entropy as a measure of spectral complexity. When the spectral entropy is lower than a preset threshold, the structure preservation mode is adopted: frequency-channel selection weights are generated through a gating network to selectively recalibrate the spectral amplitude. When the spectral entropy is higher than or equal to a preset threshold, a learnable transformation mode is adopted: the real and imaginary parts of the spectrum are respectively learned and mapped, and then recombined into a new complex frequency domain representation. The frequency domain representation is mapped back to the time domain through inverse Fourier transform, and then the frequency domain viewpoint features are output after normalization and learnable gated injection.

5. The multi-view parallel diffusion time series generation method according to claim 1, characterized in that: The multi-view parallel fusion in step S5 adopts a gated residual mechanism, which adaptively adjusts the contribution ratio of each view information through learnable gate parameters to obtain the fused multi-view encoded representation.

6. The multi-view parallel diffusion time series generation method according to claim 1, characterized in that: The decoding end uses self-attention and cross-attention joint modeling, and combines trend term and seasonal term decomposition to refine the sequence representation of the current diffusion step, outputs the original sequence estimate, further derives the state of the previous time step based on the estimate, and repeats steps S2 to S5 until all back diffusion iterations are completed, finally obtaining the generated multivariate long-term time series sample.

7. The multi-view parallel diffusion time series generation method according to claim 1, characterized in that: The time-view backbone encoder in step S2 adopts a Transformer coding structure. The input is a noisy sequence and time step embedding information, and the output is basic temporal features.

8. The multi-view parallel diffusion time series generation method according to claim 1, characterized in that: In the fluctuation sensing and control mechanism, the global state vector simultaneously encodes the overall level and fluctuation intensity of the current sample; the dynamic intensity coefficient reflects the credibility of the current dynamic graph.

9. An electronic device comprising a processor and a memory communicatively connected to the processor and used for storing processor-executable instructions, characterized in that: The processor is used to execute the method described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method described in any one of claims 1-8.

Images