Time sequence data generation method and device, electronic equipment, medium and program product

Abstract

The invention provides a time sequence data generation method and device, electronic equipment, a medium and a program product, relates to the technical field of information, and is used for improving the stability and accuracy of time sequence data generation. The method comprises the following steps: inputting a text cue word into a text time sequence diffusion model; performing coding processing on the text word segmentation sequence corresponding to the text prompt word through a text coding module of the text time sequence diffusion model to obtain a text condition vector; sampling the text condition vector and the first random noise through a scale coding module to generate a target scale latent variable; a denoising module of the text time sequence diffusion model is used for conducting denoising processing on the target scale latent variable, the processed target scale latent variable is input into a scale decoding module for decoding reconstruction processing, multi-dimensional time sequence data are generated, and the multi-dimensional time sequence data comprise statistical characteristics and time characteristics of text cues; and a dependency relationship between the statistical characteristic and the temporal characteristic. The method and the device are applied to a text time sequence scene.

Description

Technical Field

[0001] This application relates to the field of information technology, and in particular to time-series data generation methods, apparatus, electronic devices, media, and program products. Background Technology

[0002] Time series data is a collection of data recorded in chronological order, with each data point associated with a specific timestamp. Currently, text prompts can be processed using textual time series models to generate corresponding time series data. For example, inputting the prompt "stocks first rise and then fluctuate" can generate corresponding stock price trend data.

[0003] In related technologies, Wensheng time series models are mainly built based on diffusion models, which primarily include a temporal autoencoder (VAE) used to learn latent space representations to generate data in the latent space. However, existing VAEs only support univariate time series generation, leading to unstable long dependency capture and blurred details, resulting in trend errors or missing details in the generated results, thus causing poor accuracy of the generated time series data. Summary of the Invention

[0004] This application provides a time-series data generation method, apparatus, electronic device, storage medium, and program product to improve the accuracy of generated time-series data.

[0005] In a first aspect, this application provides a method for generating time-series data, comprising: inputting text prompt words into a text-generated time-series diffusion model, the text-generated time-series diffusion model comprising at least a dual latent space variational autoencoder, the dual latent space variational autoencoder comprising a text encoding module, a scale encoding module, a noise module, and a scale decoding module; encoding the text segmentation sequence corresponding to the text prompt words through the text encoding module to obtain a text condition vector; sampling the text condition vector and a first random noise through the scale encoding module to generate a target scale latent variable, the target scale latent variable representing the feature information of the time-series features contained in the text prompt words; denoising the target scale latent variable through the denoising module of the text-generated time-series diffusion model, and inputting the processed target scale latent variable into the scale decoding module for decoding and reconstruction to generate multidimensional time-series data, the multidimensional time-series data containing the statistical characteristics and temporal characteristics of the text prompt words, as well as the dependency relationship between the statistical characteristics and the temporal characteristics.

[0006] The technical solution provided in this application offers at least the following benefits: First, the user-inputted text prompts are converted into text conditional vectors; then, a scale encoding module generates target-scale latent variables representing temporal features based on the text conditional vectors and random noise; finally, a decoder reconstructs multidimensional time-series data that conforms to the text description. This solution, by processing the low-frequency trends and high-frequency details of the time-series data separately, avoids trend drift and detail distortion caused by feature confusion in a single latent space, thus ensuring that the final generated data maintains a high degree of consistency between long-term trends and short-term fluctuations, significantly improving the accuracy of the generated time-series data.

[0007] One possible implementation is that the above-mentioned scale encoding module includes a first scale encoding module, a second scale encoding module, and a fusion module; the step of sampling the text conditional vector and the first random noise through the scale encoding module to generate a target scale latent variable includes: performing flow matching sampling on the text conditional vector and the first random noise through the first scale encoding module to generate a first scale latent variable, wherein the first scale latent variable represents the low-frequency trend features of the temporal features contained in the text prompt word; performing conditional flow matching sampling on the text conditional vector, the first scale latent variable, and the second random noise through the second scale encoding module to generate a second scale latent variable, wherein the second scale latent variable represents the high-frequency detail features of the temporal features contained in the text prompt word; and synthesizing the first scale latent variable and the second scale latent variable through the fusion module to obtain the target scale latent variable.

[0008] Another possible implementation involves inputting the processed target-scale latent variables into the scale decoding module for decoding and reconstruction to generate multidimensional time-series data. This includes: decoding the target-scale latent variables through the scale decoding module to reconstruct low-frequency trend components and high-frequency detail components; wherein the low-frequency trend components contain trend feature information of the time-series features of the text prompt words, and the high-frequency detail components contain detail feature information of the time-series features of the text prompt words; and synthesizing the low-frequency trend components and the high-frequency detail components to obtain the multidimensional time-series data.

[0009] Another possible implementation, as described above, involves using the first scale encoding module to perform flow matching sampling on the text condition vector and the first random noise to generate first scale latent variables. This includes: combining the text condition vector and the first random noise using the first scale encoding module to obtain an initial latent variable sequence; segmenting the initial latent variable sequence along its dimensional direction using the first scale encoding module to obtain a dimensional sequence, and rearranging the dimensional sequence into a time window sequence; processing the dimensional sequence using the first diffusion transformer of the first scale encoding module to obtain the collaborative relationships between different dimensions; processing the time window sequence using the second diffusion transformer of the first scale encoding module to obtain the time-dimensional dependencies; and modeling based on the obtained dimensional collaborative relationships and time dependencies, and gradually denoising through flow matching sampling to generate the first scale latent variables.

[0010] Another possible implementation involves performing conditional flow matching sampling on the text conditional vector, the first scale latent variable, and the second random noise using the second scale encoding module to generate the second scale latent variable. This includes: performing feature transformation on the first scale latent variable using the condition extraction network of the second scale encoding module to generate a trend conditional vector; concatenating the trend conditional vector with the text conditional vector to obtain an enhanced conditional vector; and performing multi-step iterative denoising processing based on flow matching on the second random noise based on the enhanced conditional vector to generate the second scale latent variable corresponding to the first scale latent variable.

[0011] Another possible implementation, the above-mentioned textual sequential diffusion model further includes: a word segmentation module; the step of encoding the text segmentation sequence corresponding to the text prompt word through the text encoding module to obtain the text condition vector includes: performing word segmentation on the text prompt word through the word segmentation module to obtain the text segmentation sequence; inputting the text segmentation sequence into the text encoding module, and extracting text semantic features through the multi-layer self-attention mechanism of the text encoding module to obtain the text condition vector.

[0012] Another possible implementation, the above method further includes: performing wavelet decomposition on the time-series training data to obtain low-frequency trend components and high-frequency detail components; processing the low-frequency trend components through the first-scale encoding module of the dual-latent spatial variational autoencoder to obtain first-scale latent variable distribution parameters, and processing the low-frequency trend components through the second-scale encoding module of the dual-latent spatial variational autoencoder to obtain second-scale latent variable distribution parameters; obtaining first-scale training latent variables and second-scale training latent variables based on the first-scale latent variable distribution parameters and the second-scale latent variable distribution parameters, respectively; performing decoding and reconstruction processing on the first-scale training latent variables and the second-scale training latent variables through the scale decoding module of the dual-latent spatial variational autoencoder to obtain low-frequency trend components and high-frequency trend components, and calculating reconstruction loss; updating the parameters of the dual-latent spatial variational autoencoder based on the reconstruction loss.

[0013] Another possible implementation, after updating the parameters of the dual latent space variational autoencoder based on the reconstruction loss, further includes: freezing the updated parameters of the dual latent space variational autoencoder; constructing multidimensional temporal latent variable training samples based on the first-scale training latent variables and the second-scale training latent variables; training the first-scale encoding module and the second-scale encoding module based on the multidimensional temporal latent variable training samples, and updating the parameters of the first-scale encoding module and the second-scale encoding module respectively using a flow matching loss function.

[0014] Secondly, this application provides a time-series data generation apparatus, comprising: a processing module, the processing module being configured to: input text prompt words into a text-generated time-series diffusion model, the text-generated time-series diffusion model comprising at least: a dual latent space variational autoencoder, the dual latent space variational autoencoder comprising: a text encoding module, a scale encoding module, a noise module, and a scale decoding module; encode the text segmentation sequence corresponding to the text prompt words through the text encoding module to obtain a text condition vector; sample the text condition vector and a first random noise through the scale encoding module to generate a target scale latent variable, the target scale latent variable representing the feature information of the time-series features contained in the text prompt words; denoise the target scale latent variable through the denoising module of the text-generated time-series diffusion model, and input the processed target scale latent variable into the scale decoding module for decoding and reconstruction processing to generate multidimensional time-series data, the multidimensional time-series data containing the statistical characteristics and temporal characteristics of the text prompt words, as well as the dependency relationship between the statistical characteristics and the temporal characteristics.

[0015] In some possible embodiments, the above-mentioned scale encoding module includes a first scale encoding module, a second scale encoding module, and a fusion module; the above-mentioned processing module is specifically used for: performing flow matching sampling processing on the text conditional vector and the first random noise through the first scale encoding module to generate a first scale latent variable, wherein the first scale latent variable represents the low-frequency trend features of the temporal features contained in the text prompt word; performing conditional flow matching sampling processing on the text conditional vector, the first scale latent variable, and the second random noise through the second scale encoding module to generate a second scale latent variable, wherein the second scale latent variable represents the high-frequency detail features of the temporal features contained in the text prompt word; and performing synthesis processing on the first scale latent variable and the second scale latent variable through the fusion module to obtain the target scale latent variable.

[0016] In some other possible embodiments, the above processing module is specifically used to: decode the target scale latent variable through the scale decoding module to reconstruct low-frequency trend components and high-frequency detail components; wherein, the low-frequency trend components contain trend feature information of the temporal features of the text prompt words, and the high-frequency detail components contain detail feature information of the temporal features of the text prompt words; and synthesize the low-frequency trend components and the high-frequency detail components to obtain the multidimensional time series data.

[0017] In some other possible embodiments, the above processing module is specifically used to: combine the text condition vector with the first random noise through the first scale encoding module to obtain an initial latent variable sequence; segment the initial latent variable sequence along its dimensional direction through the first scale encoding module to obtain a dimensional sequence, and rearrange the dimensional sequence into a time window sequence; process the dimensional sequence through the first diffusion transformer of the first scale encoding module to obtain the collaborative relationship between different dimensions; process the time window sequence through the second diffusion transformer of the first scale encoding module to obtain the dependency relationship in the time dimension; and generate the first scale latent variable by modeling the obtained dimensional collaborative relationship and time dependency relationship and gradually denoising through stream matching sampling.

[0018] In some other possible embodiments, the above processing module is specifically used to: perform feature transformation on the first scale latent variable through the condition extraction network of the second scale encoding module to generate a trend condition vector; concatenate the trend condition vector with the text condition vector to obtain an enhanced condition vector; and perform multi-step iterative denoising processing based on flow matching on the second random noise based on the enhanced condition vector to generate a second scale latent variable corresponding to the first scale latent variable.

[0019] In some other possible embodiments, the textual sequential diffusion model further includes: a word segmentation module; the above processing module is specifically used to: perform word segmentation processing on the text prompt words through the word segmentation module to obtain the text word segmentation sequence; input the text word segmentation sequence into the text encoding module, and extract text semantic features through the multi-layer self-attention mechanism of the text encoding module to obtain the text condition vector.

[0020] In other possible embodiments, the above processing module is specifically used for: performing wavelet decomposition on the time-series training data to obtain low-frequency trend components and high-frequency detail components; processing the low-frequency trend components through the first scale encoding module of the dual-latent spatial variational autoencoder to obtain first-scale latent variable distribution parameters, and processing the low-frequency trend components through the second scale encoding module of the dual-latent spatial variational autoencoder to obtain second-scale latent variable distribution parameters; obtaining first-scale training latent variables and second-scale training latent variables based on the first-scale latent variable distribution parameters and the second-scale latent variable distribution parameters, respectively; performing decoding and reconstruction processing on the first-scale training latent variables and the second-scale training latent variables through the scale decoding module of the dual-latent spatial variational autoencoder to obtain low-frequency trend components and high-frequency trend components, and calculating reconstruction loss; and updating the parameters of the dual-latent spatial variational autoencoder based on the reconstruction loss.

[0021] In some other possible embodiments, the above processing module is further configured to: freeze the updated parameters of the dual latent spatial variational autoencoder; construct multidimensional temporal latent variable training samples based on the first-scale training latent variables and the second-scale training latent variables; train the first-scale encoding module and the second-scale encoding module based on the multidimensional temporal latent variable training samples, and update the parameters of the first-scale encoding module and the second-scale encoding module respectively using a flow matching loss function.

[0022] Thirdly, this application provides an electronic device comprising: a processor and a memory; the memory storing processor-executable instructions; when the processor is configured to execute the instructions, causing the electronic device to implement the method of the first aspect described above.

[0023] Fourthly, this application provides a computer-readable storage medium comprising: computer software instructions; which, when executed in an electronic device, cause the electronic device to implement the method described in the first aspect.

[0024] Fifthly, this application provides a computer program product comprising a computer program; when the computer program is run in an electronic device, the electronic device performs the method described in the first aspect.

[0025] The beneficial effects of the second to fifth aspects mentioned above are described in the corresponding description of the first aspect and will not be repeated here. Attached Figure Description

[0026] Figure 1 A schematic diagram illustrating the application environment of a time-series data generation method provided in this application; Figure 2 A flowchart illustrating a time-series data generation method provided in this application; Figure 3 A schematic diagram illustrating a time-series data generation process provided in this application; Figure 4 A schematic diagram of a VAE training process provided in this application; Figure 5 A schematic diagram illustrating the training process of a denoising module provided in this application; Figure 6 A flowchart illustrating another time-series data generation method provided in this application; Figure 7 This application provides a schematic diagram of a process for generating time-series data using a dual-latency VAE. Figure 8 A schematic diagram illustrating the processing flow of a time-dimension alternation modeling module provided in an embodiment of this application; Figure 9 A schematic diagram illustrating the processing flow of a long-scale time-alternating correlation module provided in an embodiment of this application; Figure 10 A schematic diagram illustrating the processing flow of a short-scale time-alternating correlation module provided in an embodiment of this application; Figure 11 A schematic diagram of a two-stage timing generation process provided in an embodiment of this application; Figure 12 A flowchart illustrating another time-series data generation method provided in this application; Figure 13 A schematic diagram of the system architecture of a time series data generation system provided in this application; Figure 14 A schematic diagram of the composition of a time-series data generation device provided in this application; Figure 15 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0027] The timing data generation method, apparatus, electronic device, medium, and program product provided in this application will now be described in detail with reference to the accompanying drawings.

[0028] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0029] The terms "first" and "second," etc., used in the specification and drawings of this application are used to distinguish different objects or to distinguish different treatments of the same object, rather than to describe a specific order of objects.

[0030] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0031] It should be noted that in the embodiments of this application, the words "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0032] To facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish the same or similar items with essentially the same function and effect. Those skilled in the art can understand that the terms "first" and "second" are not intended to limit the quantity or execution order.

[0033] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0034] The time-series data generation method provided in this application can be applied to scenarios involving the generation of time-series data.

[0035] For example, the embodiments of this application can be applied to the field of environmental science, where it is necessary to simulate changes in environmental parameters under different climatic conditions. For instance, the input description might be "Generate environmental monitoring data for a coastal area affected by a typhoon over the next 72 hours, including wind speed gradually increasing from 5 m / s to 25 m / s over 24 hours and maintaining this for 12 hours, air pressure synchronously decreasing from 1013 hPa to 980 hPa, and rainfall reaching 50 mm / h during peak wind speed." In related technologies, traditional numerical weather prediction models are computationally intensive and have slow response times; existing statistical methods struggle to generate multivariate collaborative data that conforms to complex meteorological conditions and cannot guarantee the physical correlation between various environmental parameters. Thus, existing technologies suffer from the problem of failing to quickly respond to specific meteorological condition descriptions and generate physically consistent multidimensional environmental monitoring data. This method, through a text-driven generation approach, can quickly generate synthetic data that conforms to specific meteorological scene descriptions and maintains reasonable physical correlations between parameters, providing support for testing environmental monitoring algorithms.

[0036] To address the aforementioned technical problems, this application provides a time-series data generation method, apparatus, electronic device, medium, and program product.

[0037] The embodiments provided in this application will now be described in detail with reference to the accompanying drawings.

[0038] The time-series data generation method provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes a time-series data generation device 101 and a front-end device 102. The time-series data generation device 101 and the front-end device 102 are interconnected.

[0039] In some embodiments, the time-series data generation device 101 may be a server cluster consisting of multiple servers, a single server, a computer, or a processor or processing chip in a server or computer, etc. This application does not limit the specific device form of the time-series data generation device 101. Figure 1 The example shown is a time-series data generation device 101, which is a single server.

[0040] In some embodiments, the front-end device 102 can be a device with wireless transceiver capabilities, such as a mobile phone, tablet computer, wearable device, in-vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), etc. This application embodiment does not limit the specific device form of the front-end device 102. Figure 1 The example shown is a mobile phone, with the front-end device 102 as the illustration.

[0041] In some embodiments, the time series data generation device 101 inputs text prompt words into the Wensheng time series diffusion model; the text encoding module of the Wensheng time series diffusion model encodes the text segmentation sequence corresponding to the text prompt words to obtain a text condition vector; the scale encoding module samples the text condition vector and the first random noise to generate a target scale latent variable; the denoising module of the Wensheng time series diffusion model denoises the target scale latent variable, and the processed target scale latent variable is input into the scale decoding module for decoding and reconstruction to generate multidimensional time series data, and then the multidimensional time series data is sent to the front-end device 102.

[0042] It should be noted that the system architecture described in the embodiments of this application is for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and does not constitute a limitation on the technical solutions provided in the embodiments of this application. As those skilled in the art will know, with the evolution of system architecture, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0043] See Figure 2 This is a flowchart illustrating a time-series data generation method provided in an embodiment of this application. Figure 2 As shown, the time series data generation method provided in this application can be implemented by the aforementioned time series data generation device, specifically including the following steps 201 to 204: Step 201: The time series data generation device inputs the text prompt words into the text generation time series diffusion model.

[0044] The aforementioned textual time-series diffusion model includes at least a dual latent space variational autoencoder, which comprises a text encoding module, a scale encoding module, a noise module, and a scale decoding module.

[0045] In some embodiments of this application, the aforementioned text prompts are time-series data generation requirements expressed by users in natural language.

[0046] In some embodiments of this application, the above-described textual time-series diffusion model is a deep learning architecture based on diffusion models, specifically designed to convert textual descriptions into corresponding multidimensional time-series data.

[0047] In some embodiments of this application, the core component of the above-mentioned textual temporal diffusion model is a dual latent space variational autoencoder, which includes at least a text encoding module, a scale encoding module, a noise module, and a scale decoding module.

[0048] Specifically, text prompts are input into the model after preprocessing. The text encoding module is responsible for mapping text information into numerical feature representations; the scaling encoding module is used to extract temporal features at different time scales; the noise module introduces randomness to increase the diversity of generated data; and the scaling decoding module reconstructs latent variables into the final multidimensional time-series data.

[0049] Step 202: The time-series data generation device encodes the text segmentation sequence corresponding to the text prompt word through the text encoding module to obtain the text condition vector.

[0050] In some embodiments of this application, the input text prompts are first segmented using a text encoding module, converting them into a text segmentation sequence. Then, a multi-layer neural network structure is used to encode the segmentation sequence and extract deep semantic features.

[0051] Specifically, the text encoding module adopts a Transformer-based architecture and uses a self-attention mechanism to capture long-distance dependencies in the text.

[0052] Step 203: The time series data generation device samples the text condition vector and the first random noise through the scale coding module to generate target scale latent variables.

[0053] Among them, the aforementioned target-scale latent variables represent the temporal features contained in the text prompt words.

[0054] In some embodiments of this application, the scale encoding module receives a text conditional vector and a first random noise as input, and generates a target scale latent variable through sampling processing. The target scale latent variable is a compact numerical representation that encodes the temporal feature information contained in the text description.

[0055] Specifically, the scaling module employs variational inference to learn the mapping relationship between the text conditional vectors and the latent space. The sampling process is based on reparameterization techniques, allowing gradients to propagate back through random nodes.

[0056] Step 204: The time series data generation device uses the denoising module of the Wensheng time series diffusion model to denoise the target scale latent variables, and inputs the denoised target scale latent variables into the scale decoding module for decoding and reconstruction to generate multi-dimensional time series data.

[0057] The aforementioned multidimensional time-series data includes the statistical and temporal characteristics of text prompt words, as well as the dependencies between these statistical and temporal characteristics.

[0058] In some embodiments of this application, the target-scale latent variables are refined by a denoising module to remove noise components and improve generation quality. The processed target-scale latent variables are then input into a scale decoding module and reconstructed into multidimensional time-series data through a multi-layer neural network.

[0059] In some embodiments of this application, the scaling decoding module employs a deconvolutional network or a Transformer decoder architecture to progressively upsample latent variables to the target temporal length.

[0060] For example, in an industrial equipment monitoring scenario, the input text description is "to generate multi-dimensional time-series data of a centrifugal compressor from a healthy state to a state of slight bearing wear, during which the vibration signal slowly increases by 20% over 100 hours". The time-series data generation device converts the description into a conditional vector through a text encoding module, then generates a scale latent variable containing trend information through a scale encoding module, and finally outputs the equipment vibration time-series data that meets the requirements through a scale decoding module.

[0061] The time-series data generation method provided in this application can complete the modeling and generation of time-series data within the latent space. The following provides an exemplary description of the textual diffusion model of time series data provided in this application.

[0062] The following provides an exemplary description of the time series data generation process of the Wensheng time series diffusion model.

[0063] For example, such as Figure 3 As shown, noise is first sampled from a standard Gaussian distribution, forming an initial noise latent variable, which serves as the starting point for the generation process. This noise latent variable and the encoded text description are then input into a trained denoising module, which updates the latent variable progressively from large to small over time, moving towards the target distribution along the learned probability flow at each step. After multiple iterations, the noise component is gradually removed, ultimately yielding a target latent variable consistent with the training data distribution. This target latent variable is then fed into a VAE decoder for decoding, reconstructing a complete multidimensional time-series data output, i.e., a time-series signal. In this way, this application can automatically generate high-fidelity time-series samples with the statistical characteristics and time dependencies described in the text, starting from pure noise, and can achieve diverse generation results by changing the initial noise.

[0064] The training process of the variational autoencoder (VAE) and denoising module of the Wensheng temporal diffusion model is illustrated below.

[0065] For example, this paper first learns the latent space representation of the original multidimensional time series data through a variational autoencoder (VAE) to enable efficient modeling and subsequent generation of time series data in a low-dimensional latent space. Figure 4 As shown, during the training process of VAE, multidimensional time-series data is first input into the VAE encoder as training samples for feature extraction, mapping the original high-dimensional time-series data to the distribution parameters (mean and variance) of the latent space, i.e., the latent variable distribution. Then, the latent variables are sampled using a resampling method to obtain the latent space representation. Subsequently, the sampled latent variables are reconstructed using a VAE decoder to obtain the reconstructed time series signal. The loss between the original time series and the reconstructed time series (i.e., reconstruction loss) is calculated, and backpropagation is performed in conjunction with the KL divergence loss to complete the learning of the latent space, ensuring that the latent variable distribution is smooth and reversible.

[0066] For example, after obtaining a stable latent space, the parameters of the VAE are frozen, and the denoising module is trained. The denoising module mainly consists of a Diffusion Transformer (DiT) and uses flow matching to learn a continuous denoising path from the noise distribution to the target distribution in the latent space. Specifically, as... Figure 5 As shown, initial noisy latent variables are obtained by sampling from a standard Gaussian distribution and adding noise. A random continuous time step is then selected, and this noisy latent variable is weighted and mixed with the true latent variable according to the proportion of that time step, resulting in a noisy latent variable in an intermediate state. This noisy latent variable, time step information, and text description encoded by a VAE encoder are input into the denoising module. The module outputs a prediction of the direction of change of the latent variable, indicating how to move the noisy latent variable closer to the true latent variable at that time step. By minimizing the loss (i.e., error) between the prediction result and the true direction of change, a flow field model that accurately describes the gradual transition of the latent variable from noise to the target distribution over time is trained. In this way, the denoising module can learn how to gradually remove noise and recover the target latent variable along the correct trajectory in the latent space.

[0067] The time-series data generation method provided in this application first converts user-inputted text prompts into text conditional vectors; then, through a scaling encoding module, it generates target-scale latent variables representing time-series features based on the text conditional vectors and random noise; finally, it reconstructs multi-dimensional time-series data that conforms to the text description through a decoder. This approach, by processing low-frequency trends and high-frequency details of the time-series data separately, avoids trend drift and detail distortion caused by feature confusion in a single latent space, thereby ensuring that the final generated data maintains a high degree of consistency between long-term trends and short-term fluctuations, significantly improving the accuracy of the generated time-series data.

[0068] In some embodiments of this application, the scale encoding module includes a first scale encoding module, a second scale encoding module, and a fusion module; exemplarily, in conjunction with the above... Figure 2 ,like Figure 6 As shown, step 203 above may include steps 203a to 203c: Step 203a: The time series data generation device performs stream matching sampling processing on the text condition vector and the first random noise through the first scale encoding module to generate the first scale latent variable.

[0069] Among them, the aforementioned first-scale latent variable represents the low-frequency trend characteristics of the temporal features contained in the text prompt words.

[0070] Step 203b: The time series data generation device performs conditional flow matching sampling processing on the text conditional vector, the first scale latent variable and the second random noise through the second scale encoding module to generate the second scale latent variable.

[0071] Among them, the aforementioned second-scale latent variables characterize the high-frequency detailed features of the temporal features contained in the text prompt words.

[0072] Step 203c: The time series data generation device synthesizes the first-scale latent variables and the second-scale latent variables through the fusion module to obtain the target-scale latent variables.

[0073] In some embodiments of this application, the first-scale encoding module described above is specifically responsible for extracting low-frequency trend features from the text description.

[0074] In some embodiments of this application, the first-scale encoding module employs a flow-matching sampling technique to generate a first-scale latent variable from the text condition vector and the first random noise by solving the probability flow ordinary differential equation.

[0075] Specifically, the stream matching sampling process is defined as: dz / dt = v(z,t,C), where v is the learned vector field, t is the time step, and C is the text condition vector. By numerically solving this equation, the first-scale latent variable z_l, representing the low-frequency trend, is obtained.

[0076] In some embodiments of this application, the time-series data generation device generates second-scale latent variables characterizing high-frequency details by using a second-scale encoding module under the condition of first-scale latent variables.

[0077] It should be noted that this second-scale encoding module uses conditional flow matching sampling to ensure that the generated detailed features are consistent with the trend features.

[0078] For example, the conditional sampling process is represented as: dz / dt = v(z,t,C,z_l), where z_l is a first-scale latent variable that serves as conditional information to guide the sampling process.

[0079] In some embodiments of this application, the time series data generation device synthesizes the first-scale latent variable and the second-scale latent variable through a fusion module to obtain the complete target-scale latent variable.

[0080] For example, the synthesis of first-scale latent variables and second-scale latent variables can be achieved through splicing, weighted summation, or attention-based fusion mechanisms.

[0081] For example, in a medical electrocardiogram (ECG) generation scenario, a basic heart rhythm trend is generated through a first-scale encoding module, and specific detailed features such as P waves and QRS complexes are generated based on the trend through a second-scale encoding module. Then, a fusion module fuses the latent variables generated by the first-scale encoding module and the second-scale encoding module to obtain a complete ECG signal.

[0082] The following provides an exemplary description of the variational autoencoder based on dual latent space provided in the embodiments of this application.

[0083] For example, such as Figure 7 As shown, the input time-series data is first fed into a shared VAE encoder, and a high-dimensional feature representation is obtained through a unified feature extraction network. At the end of the encoder, two latent variable output heads are designed to predict the distribution parameters of long-scale latent variables and short-scale latent variables, respectively, i.e., the distribution parameters of the long-scale latent variables. and short-scale latent variable distribution Next, the distribution of long-scale latent variables was analyzed. and short-scale latent variable distribution Resampling is performed, and the sampled latent variables are input into the corresponding long-scale VAE decoder and short-scale VAE decoder, respectively. The corresponding long-scale components are obtained. and short-scale components and for long-scale components and short-scale components Perform timing synthesis and output timing sequence. This allows for the simultaneous acquisition of long-scale latent variables describing global trends and short-scale latent variables capturing local details within the same encoder framework, giving the latent space a natural multi-level expressive capability.

[0084] For example, to ensure the separation of responsibilities between long and short latent variables, this application introduces a multi-scale supervision mechanism based on wavelet decomposition during the training phase. Specifically, the original time-series signal is decomposed using wavelet decomposition to obtain low-frequency and high-frequency sub-band signals. Long-scale latent variables reconstruct the low-frequency sub-band signal using a long-scale decoder, while short-scale latent variables reconstruct the high-frequency sub-band signal using a short-scale decoder. The low-frequency and high-frequency sub-band signals are then reconstructed through wavelet inverse synthesis or direct superposition to rebuild the time-series signal. During training, the low-frequency reconstruction loss, high-frequency reconstruction loss, and time-series signal reconstruction loss are calculated separately, and the KL divergence loss of the long-scale and short-scale latent variable distributions is jointly optimized. This guides the long-scale latent variables to focus on learning the global smooth trend, while the short-scale latent variables focus on modeling transient details. This multi-scale supervision strategy effectively avoids the aliasing of information between long and short latent variables, improves the interpretability of the latent space and the physical consistency of the generated results, thereby solving the common trend drift and detail ambiguity problems in traditional single-latent-space VAEs for long sequence modeling, and significantly improving the stability and clarity of the generated signal.

[0085] In this embodiment, decoupling of trends and details is achieved through scaled encoding, avoiding feature confusion. Conditional generation ensures the coordination between features at different scales, thereby improving the flexibility and scalability of the Wensheng time-series diffusion model.

[0086] In some embodiments of this application, the process of inputting the processed target-scale latent variables into the scale decoding module for decoding and reconstruction in step 204 above to generate multidimensional time series data may include the following steps 204a and 204b: Step 204a: The time series data generation device decodes the target scale latent variable through the scale decoding module to reconstruct the low-frequency trend component and the high-frequency detail component.

[0087] Among them, the low-frequency trend component contains the trend feature information of the temporal features of the text prompt words, and the high-frequency detail component contains the detail feature information of the temporal features of the text prompt words.

[0088] Step 204b: The time series data generation device synthesizes the low-frequency trend component and the high-frequency detail component to obtain multi-dimensional time series data.

[0089] In some embodiments of this application, the scale decoding module described above includes dedicated branches for reconstructing time-series components of different scales, namely, low-frequency decoding branches and high-frequency decoding branches.

[0090] In some embodiments of this application, the low-frequency decoding branch reconstructs the trend component through a fully connected layer and a temporal convolutional network, while the high-frequency decoding branch reconstructs the detail component through a more refined deconvolutional network.

[0091] In some embodiments of this application, the time series data generation device reconstructs low-frequency trend components and high-frequency detail components, and synthesizes the reconstructed low-frequency trend components and high-frequency detail components to obtain the final multi-dimensional time series data.

[0092] In some embodiments of this application, the time series data generation device may employ an adaptive weighting method to dynamically adjust the component weights according to the feature importance of different time points, so as to synthesize the reconstructed low-frequency trend components and high-frequency detail components.

[0093] For example, in robot trajectory generation, low-frequency components generate smooth motion trajectories, while high-frequency components add subtle adjustment movements, resulting in a smooth and accurate complete trajectory after synthesis.

[0094] In this embodiment, by reconstructing low-frequency trend components and high-frequency detail components, independent component reconstruction avoids information loss and maintains the integrity of features. The contribution of features at different scales is optimized through an adaptive synthesis mechanism, thereby improving the accuracy and practicality of the generated time series data.

[0095] In some embodiments of this application, step 203a may include steps A1 to A5: Step A1: The time series data generation device combines the text condition vector with the first random noise through the first scale encoding module to obtain the initial latent variable sequence.

[0096] Step A2: The time series data generation device divides the initial latent variable sequence along its dimensional direction through the first scale encoding module to obtain the dimensional sequence, and rearranges the dimensional sequence into a time window sequence; Step A3: The time series data generation device processes the dimensional sequence through the first diffusion transformer of the first scale encoding module to obtain the collaborative relationship between different dimensions; Step A4: The time series data generation device processes the time window sequence through the second diffusion transformer of the first scale encoding module to obtain the dependencies in the time dimension; Step A5: The time series data generation device models the acquired dimensional collaborative relationship and time dependency relationship, and gradually denoises through stream matching sampling to generate the first-scale latent variable.

[0097] In some embodiments of this application, the time-series data generation device combines the text condition vector with the first random noise through the first scale encoding module to form an initial latent variable sequence.

[0098] For example, the time series data generation device can use the first scale encoding module to concatenate the text condition vector with the vector corresponding to the first random noise to obtain the initial latent variable sequence.

[0099] In some embodiments of this application, the time series data generation device can divide the initial latent variable sequence into multiple dimensional sequences along the dimensional direction, and then rearrange the dimensional sequences into a time window sequence.

[0100] In some embodiments of this application, the time-series data generation device processes the dimensional sequence through a first diffusion transformer and learns the collaborative relationship between different dimensions through a self-attention mechanism.

[0101] In some embodiments of this application, the time-series data generation device processes the time window sequence through a second diffusion transformer to learn the dependencies in the time dimension. A causal attention mechanism is employed to ensure the temporal dependencies.

[0102] In some embodiments of this application, the time-series data generation device generates first-scale latent variables by gradually denoising through stream matching sampling based on learned dimensional collaboration relationships and time dependencies.

[0103] For example, in climate data generation, the dimension transformer learns the correlation between dimensions such as temperature, humidity, and air pressure, while the time transformer learns the temporal evolution of meteorological elements, ultimately generating a consistent multidimensional climate sequence.

[0104] It should be noted that most existing Wensheng time series generation methods only support the generation of single-dimensional time series data, lacking explicit modeling of the correlation between multi-dimensional time series, and cannot guarantee the synchronization and coordination characteristics between signals of different dimensions. For example, in multi-lead electrocardiograms and other multi-dimensional physiological signal scenarios, the waveforms of different leads have obvious temporal correspondences. Existing methods often cannot simultaneously maintain the waveform morphology and timing consistency of each channel, easily leading to problems such as cross-channel misalignment and amplitude mismatch.

[0105] In this embodiment, a multi-dimensional time series generation task is achieved using a dimensional time alternation modeling method and a diffusion transformer (DiT). End-to-end joint modeling and generation of multi-channel time series signals is realized through a dimensional time alternation correlation unit. This method can simultaneously learn the cooperative relationships between dimensions and the dynamic dependencies of the time series, ensuring that the generated multi-dimensional signals remain synchronized, coherent, and physically reasonable across channels. Specifically, as... Figure 8As shown, in the dimensional-time alternation correlation unit, a batch of data (labeled as Nx) of multidimensional time series is first divided into independent token sequences along the dimensional direction. The correlation and cooperation patterns between dimensions are learned through DiT, i.e., the correlation between the text latent vector and time t. Then, the output is transformed and rearranged into a sequence arranged according to time and dimension, i.e., a time window token sequence. The dynamic dependency relationship in the time direction is learned through DiT, i.e., the correlation between the text latent vector and time t, and the output is transformed again. This constitutes the dimensional-time alternation correlation unit, and multiple dimensional-time alternation correlation units together constitute the dimensional-time alternation modeling module. The above dimensional modeling and time modeling are performed alternately, iterating multiple times until convergence, allowing the model to gradually optimize the cooperation between dimensions and the smoothness of the time series, thereby generating multidimensional time series signals that simultaneously satisfy cross-dimensional consistency and temporal rationality. This method can significantly improve the synchronization, coherence, and physical rationality of multidimensional signals. Compared with traditional single-dimensional or one-time modeling methods, the generated results are more consistent with the real distribution and are suitable for complex multi-channel scenarios such as multi-lead electrocardiograms, industrial multi-sensor signals, and robot multi-joint trajectories.

[0106] In this embodiment, the structural characteristics of multidimensional time series data are fully explored through alternating time-dimensional modeling, powerful feature extraction capabilities are provided through diffusion transformers, and the stability and efficiency of the generation process are ensured through flow matching sampling, thereby improving the accuracy and efficiency of Wensheng time series diffusion model in generating smooth data.

[0107] In some embodiments of this application, the process of generating the second-scale latent variable by performing conditionalized stream matching sampling on the text conditional vector, the first-scale latent variable, and the second random noise through the second-scale coding module in step 203a above may include the following steps B1 to B3: Step B1: The time series data generation device performs feature transformation on the first-scale latent variables through the condition extraction network of the second-scale encoding module to generate a trend condition vector.

[0108] Step B2: The time series data generation device concatenates the trend condition vector with the text condition vector to obtain the enhanced condition vector.

[0109] Step B3: The time series data generation device performs multi-step iterative denoising processing on the second random noise based on flow matching, based on the enhanced conditional vector, to generate a second-scale latent variable corresponding to the first-scale latent variable.

[0110] In some embodiments of this application, the time series data generation device uses a multilayer perceptron structure through a condition extraction network to perform nonlinear transformation on the first-scale latent variables to generate a trend condition vector.

[0111] In some embodiments of this application, the time-series data generation device concatenates the trend condition vector with the text condition vector to obtain an enhanced condition vector.

[0112] Understandably, the concatenated enhanced conditional vector contains both textual semantics and trend features.

[0113] In some embodiments of this application, the time-series data generation apparatus performs conditional flow matching sampling on the second random noise based on the enhanced conditional vector, and performs multi-step iterative denoising to generate a second-scale latent variable that is consistent with the first-scale latent variable.

[0114] For example, in speech signal generation, trend conditional vectors encode fundamental frequency profiles, text conditional vectors provide phoneme information, and conditional generation ensures the prosodic naturalness and semantic accuracy of synthesized speech.

[0115] It should be noted that current single-stage time series generation methods typically learn long-term trends and short-term details simultaneously within the same latent space. Since trends and details differ significantly in statistical characteristics and time scales, the model is prone to trade-offs during training, making it difficult to balance global consistency and local realism.

[0116] For example, this application explicitly decouples the generation process of global trends and local details through a two-stage textual time series generation method. In the training phase, the first-stage long-scale time alternation modeling module is trained on the long-scale latent space. The input sequence is segmented, and flow matching is used to optimize the vector field, which is then input into the long-scale time alternation modeling module. Additionally, the text description is segmented, processed using flow matching, and then input into the long-scale time alternation modeling module to obtain long-scale latent variables. Specifically, as... Figure 9 As shown, the input noisy data is segmented to obtain a time-series data sequence. This time-series data sequence is then input for sampling and processed using addition. Additionally, the text description is segmented, and the resulting sequence is sampled and added again. The processed sequence and time t are then input into the long-scale time-alternation correlation module (i.e., the long-scale time-alternation modeling module) to obtain long-scale latent variables. In this way, the noise distribution is smoothly transformed into the true long-scale latent variable distribution, thereby learning a stable global trend generator.

[0117] For example, after the long-scale network converges, its parameters are frozen. A trend conditional vector is generated using a conditional extraction network (such as a multilayer perceptron) based on the real long-scale latent variables. This conditional vector is then embedded into the input of the second-stage short-scale time-alternating modeling module, ensuring that it coordinates with the long-scale trend while generating short-scale latent variables, thereby learning a conditional short-scale vector field. Figure 10As shown, long-scale latent variables are input into a multilayer perceptron for processing, and the input noise data is segmented. The processed data is then concatenated with the processed long-scale latent variables, sampled, and added before being input into a short-scale time-alternating correlation module. Similarly, the data after the text description is segmented, sampled, and added is input into the short-scale time-alternating correlation module for processing to obtain short-scale latent variables.

[0118] During the generation phase, such as Figure 11 As shown, starting with random noise, the input noise data is segmented in the first stage. The processed data is then sampled and added before being input into a pre-trained long-scale time-alternating correlation module for flow matching sampling based on time t, resulting in long-scale latent variables that conform to the distribution. These latent variables are then processed and input into a condition extraction network to generate a condition vector. In the second stage, another set of input noise data is segmented and concatenated with the generated condition vector before being input into a short-scale time-alternating correlation module (i.e., the short-scale time-alternating modeling module). Text descriptions are also sampled and added before being input into the same short-scale time-alternating correlation module for flow matching sampling based on time t. Finally, the two types of latent variables output from the long-scale and short-scale time-alternating correlation modules are fed into the long-scale and short-scale VAE decoders, respectively, to reconstruct low-frequency and high-frequency components and perform time-series synthesis, resulting in a complete multi-dimensional time series sequence that has both a smooth and reasonable trend and contains detailed fluctuations.

[0119] In this embodiment, the two-stage generation strategy can significantly alleviate the conflict between trends and details, ensure the global consistency and stability of long sequences, and retain rich local details, making the generated results more consistent with the distribution of real time-series signals. It is especially suitable for complex scenarios that require both trends and details, such as multi-dimensional sensor signals, industrial monitoring data, and robot control trajectories.

[0120] In this embodiment, the first-scale latent variable is processed by a conditional extraction network to effectively transmit trend information, thereby enhancing the conditional vector to integrate multi-source information and ensuring the consistency of multi-scale features.

[0121] In some embodiments of this application, the above-described textual sequential diffusion model further includes: a word segmentation module; exemplarily, in conjunction with the above... Figure 2 ,like Figure 12 As shown, the process of step 202 above may include the following steps 202a and 202b: Step 202a: The time-series data generation device performs word segmentation on the text prompt words through the word segmentation module to obtain a text word segmentation sequence.

[0122] Step 202b: The time-series data generation device inputs the text segmentation sequence into the text encoding module, and extracts the text semantic features through the multi-layer self-attention mechanism of the text encoding module to obtain the text condition vector.

[0123] In some embodiments of this application, the time-series data generation device uses a word segmentation module to segment text prompts into meaningful lexical units using a dictionary-based or neural network-based word segmentation method.

[0124] For example, for Chinese text, BPE or WordPiece segmentation algorithms are used; for English text, space segmentation or sub-word segmentation are used.

[0125] In some embodiments of this application, the time-series data generation device uses a pre-trained language model, such as BERT or RoBERTa, through a text encoding module to extract text semantic features through a multi-layer self-attention mechanism.

[0126] For example, when generating equipment fault data, the input "early wear vibration characteristics of bearings" is segmented into ["bearing", "early", "wear", "vibration", "characteristics"], and after text encoding, a conditional vector containing complete semantics is obtained.

[0127] In this embodiment of the application, by performing word segmentation on the text prompt words and subsequent encoding on the resulting text sequence, the characteristics and semantics of the text prompt words can be obtained more accurately, thereby ensuring the processing performance of the text generation time-series diffusion model.

[0128] In some embodiments of this application, the time-series data generation method provided in this application may further include steps 205 to 209: Step 205: The time series data generation device performs wavelet decomposition on the time series training data to obtain low-frequency trend components and high-frequency detail components.

[0129] Step 206: The time series data generation device processes the low-frequency trend component through the first-scale encoding module of the dual latent spatial variational autoencoder to obtain the first-scale latent variable distribution parameters, and processes the low-frequency trend component through the second-scale encoding module of the dual latent spatial variational autoencoder to obtain the second-scale latent variable distribution parameters.

[0130] Step 207: The time series data generation device obtains the first-scale training latent variables and the second-scale training latent variables based on the first-scale latent variable distribution parameters and the second-scale latent variable distribution parameters, respectively.

[0131] Step 208: The time series data generation device decodes and reconstructs the first-scale training latent variables and the second-scale training latent variables through the scale decoding module of the dual latent spatial variational autoencoder to obtain the low-frequency trend component and the high-frequency trend component, and calculates the reconstruction loss.

[0132] Step 209: The time-series data generation device updates the parameters of the dual latent space variational autoencoder based on the reconstruction loss.

[0133] In some embodiments of this application, the time series data generation device can use discrete wavelet transform to perform wavelet decomposition on the time series training data to obtain low-frequency trend components and high-frequency detail components.

[0134] In some embodiments of this application, the time-series data generation device processes low-frequency components through a first-scale encoding module and outputs first-scale latent variable distribution parameters (mean μ_l and variance σ_l); and processes high-frequency components through a second-scale encoding module and outputs second-scale latent variable distribution parameters (mean μ_h and variance σ_h).

[0135] In some embodiments of this application, the time series data generation device obtains training latent variables by sampling through reparameterization techniques based on distribution parameters (i.e., first-scale latent variable distribution parameters and second-scale latent variable distribution parameters).

[0136] In some embodiments of this application, the time series data generation device decodes and reconstructs the training latent variables through a scale decoding module to calculate the reconstruction loss.

[0137] In some embodiments of this application, the time-series data generation device updates the parameters of the dual latent space variational autoencoder based on the calculated reconstruction loss using a backpropagation algorithm.

[0138] In this embodiment, wavelet decomposition is performed on the time-series training data to obtain low-frequency trend components and high-frequency detail components, thereby obtaining an effective multi-scale supervision signal. Then, a dual latent space variational autoencoder is trained based on the low-frequency trend components and high-frequency detail components, which can more effectively train the dual latent space variational autoencoder and thus ensure the processing performance of the trained dual latent space variational autoencoder.

[0139] In some embodiments of this application, after step 209 above, the time-series data generation method provided in this application may further include steps 210 to 212: Step 210: The time-series data generation device freezes the updated parameters of the dual latent space variational autoencoder.

[0140] Step 211: The time series data generation device constructs multidimensional time series latent variable training samples based on the first-scale training latent variables and the second-scale training latent variables.

[0141] Step 212: The time series data generation device trains the first-scale encoding module and the second-scale encoding module based on the multi-dimensional time series latent variable training samples, and updates the parameters of the first-scale encoding module and the second-scale encoding module respectively using the flow matching loss function.

[0142] In some embodiments of this application, after the time-series data generation device has completed the training of the dual latent space variational autoencoder, it freezes its parameters and trains only the first-scale coding module and the second-scale coding module.

[0143] In some embodiments of this application, the time series data generation device constructs multidimensional time series latent variable training samples based on training latent variables.

[0144] In some embodiments of this application, the time-series data generation device trains a first-scale encoding module and a second-scale encoding module based on multi-dimensional training samples and using a stream matching loss function.

[0145] For example, during the training phase, trend and detail components are obtained from electrocardiogram data through wavelet decomposition, and a dual latent space variational autoencoder is trained. Then, the parameters of the trained dual latent space variational autoencoder are frozen, and the first-scale encoding module and the second-scale encoding module are trained separately to optimize the generation quality.

[0146] In this embodiment, after training the dual latent space variational autoencoder, its parameters are frozen, and multidimensional temporal latent variable training samples are constructed. The first-scale encoding module and the second-scale encoding module are trained. The training effect of each module of the model is improved by the phased training strategy, and the stability of the generation process is improved by the flow matching loss, thereby improving the training effect of the model.

[0147] The time-series data generation method provided in this application simultaneously outputs long-scale and short-scale latent variables based on a shared encoder. It utilizes low-frequency and high-frequency components obtained from wavelet decomposition as reconstruction targets to perform multi-scale supervision of the VAE, ensuring that long-scale latent variables learn the global trend and short-scale latent variables learn local details. By reconstructing low-frequency and high-frequency components separately using dual decoders and then synthesizing them, the multi-scale latent space is explicitly decoupled, solving the problems of trend drift and detail ambiguity in single-latest-space VAEs, and improving the interpretability of the latent space and the stability, clarity, and controllability of the generated results. Furthermore, through multi-dimensional time-series alternating modeling, it utilizes a diffusion transformer (DiT) for the first time to generate multi-dimensional time-series data. The multi-dimensional time series is divided into token sequences along the dimensional direction. First, the inter-dimensional correlations are modeled, then the tokens are rearranged into time windows to learn time dependencies, and this process is iterated repeatedly until convergence. This method ensures that the generated results remain synchronized, coherent, and physically reasonable across dimensions, addressing the problem that existing textual time series generation methods cannot guarantee cross-channel consistency and time series synchronization. It is applicable to multi-channel scenarios such as multi-lead electrocardiograms, industrial multi-sensor signals, and robot multi-joint trajectories. Furthermore, it decouples the generation processes of trends and details through a two-stage time series generation approach. First, flow matching training is performed in the long-scale latent space to generate stable long-scale latent variables. Then, these variables are used as conditions and embedded into the short-scale generation module through a condition extraction network to train the short-scale vector field, generating detailed latent variables that are consistent with the trend. During generation, the trend is first generated from noise, then details are conditionally generated, and finally, the complete time series is synthesized. This two-stage generation approach, starting with the trend and then focusing on the details, avoids mutual interference between trends and details, ensuring global consistency and detail fidelity in long sequences.

[0148] Figure 13 This is a schematic diagram of a system architecture according to an embodiment of this application. The system can be a time-series data generation system. The time-series data generation system 800 may include: a data input module 801, an encoding module 802, and a noise reduction module 803.

[0149] The data input module 801 is used to input text prompts into the text generation time-series diffusion model. The text generation time-series diffusion model includes at least a dual latent space variational autoencoder. The dual latent space variational autoencoder includes a text encoding module, a scale encoding module, a noise module, and a scale decoding module, which are applied to the relevant schemes in steps 201 and 201 above.

[0150] Encoding module 802 is used to encode the text segmentation sequence corresponding to the text prompt words to obtain a text condition vector. The text condition vector and a first random noise are sampled and processed by a scaling encoding module to generate a target scale latent variable. The target scale latent variable represents the temporal feature information contained in the text prompt words. This is applied to steps 202 and 203 above, as well as related schemes of 202 and 203.

[0151] The denoising module 803 is used to denoise the target-scale latent variables and input the denoised target-scale latent variables into the scale decoding module for decoding and reconstruction to generate multi-dimensional time-series data. The multi-dimensional time-series data includes the statistical and temporal characteristics of the text prompt words, as well as the dependencies between the statistical and temporal characteristics. This is applied to steps 204 and related schemes mentioned above.

[0152] It should be noted that for a detailed explanation of the steps performed by each module and their beneficial effects, please refer to the description in the above embodiments, which will not be repeated here.

[0153] As can be seen, the above mainly describes the solutions provided by the embodiments of this application from a methodological perspective. To achieve the above functions, the embodiments of this application provide corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the modules and algorithm steps of the various examples described in the embodiments disclosed herein, the embodiments of this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0154] This application embodiment can divide the timing data generation device into functional modules according to the above method example. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. Optionally, the module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0155] In some embodiments, this application also provides a time-series data generation apparatus. This time-series data generation apparatus may include one or more functional modules for implementing the time-series data generation method of the above method embodiments.

[0156] For example, Figure 14 This is a schematic diagram illustrating the composition of a timing data generation apparatus provided in an embodiment of this application. Figure 14 As shown, the time-series data generation device 900 includes a processing module 901.

[0157] The aforementioned processing module 901 is configured to: input text prompt words into a text-based temporal diffusion model, wherein the text-based temporal diffusion model includes at least a dual latent space variational autoencoder, wherein the dual latent space variational autoencoder includes a text encoding module, a scale encoding module, a noise module, and a scale decoding module; encode the text segmentation sequence corresponding to the text prompt words through the text encoding module to obtain a text condition vector; sample the text condition vector and a first random noise through the scale encoding module to generate a target scale latent variable, wherein the target scale latent variable represents the feature information of the temporal features contained in the text prompt words; denoise the target scale latent variable through the denoising module of the text-based temporal diffusion model, and input the processed target scale latent variable into the scale decoding module for decoding and reconstruction to generate multidimensional temporal data, wherein the multidimensional temporal data includes the statistical characteristics and temporal characteristics of the text prompt words, as well as the dependency relationship between the statistical characteristics and the temporal characteristics.

[0158] The time-series data generation apparatus provided in this application first converts user-inputted text prompts into text conditional vectors; then, through a scaling encoding module, target-scale latent variables representing time-series features are generated based on the text conditional vectors and random noise; finally, a decoder reconstructs multi-dimensional time-series data that conforms to the text description. This scheme, by processing the low-frequency trends and high-frequency details of the time-series data separately, avoids trend drift and detail distortion caused by feature confusion in a single latent space, thereby ensuring that the final generated data maintains a high degree of consistency between long-term trends and short-term fluctuations, significantly improving the accuracy of the generated time-series data.

[0159] In some embodiments, the above-mentioned scaling encoding module includes a first scaling encoding module, a second scaling encoding module, and a fusion module; the above-mentioned processing module is specifically used for: performing flow matching sampling processing on the text conditional vector and the first random noise through the first scaling encoding module to generate a first scaling latent variable, wherein the first scaling latent variable represents the low-frequency trend features of the temporal features contained in the text prompt word; performing conditional flow matching sampling processing on the text conditional vector, the first scaling latent variable, and the second random noise through the second scaling encoding module to generate a second scaling latent variable, wherein the second scaling latent variable represents the high-frequency detail features of the temporal features contained in the text prompt word; and performing synthesis processing on the first scaling latent variable and the second scaling latent variable through the fusion module to obtain the target scaling latent variable.

[0160] In other embodiments, the above processing module is specifically used to: decode the target scale latent variable through the scale decoding module to reconstruct low-frequency trend components and high-frequency detail components; wherein, the low-frequency trend component contains trend feature information of the temporal features of the text prompt word, and the high-frequency detail component contains detail feature information of the temporal features of the text prompt word; and synthesize the low-frequency trend component and the high-frequency detail component to obtain the multidimensional time series data.

[0161] In other embodiments, the above processing module is specifically used to: combine the text condition vector with the first random noise through the first scale encoding module to obtain an initial latent variable sequence; segment the initial latent variable sequence along its dimensional direction through the first scale encoding module to obtain a dimensional sequence, and rearrange the dimensional sequence into a time window sequence; process the dimensional sequence through the first diffusion transformer of the first scale encoding module to obtain the collaborative relationship between different dimensions; process the time window sequence through the second diffusion transformer of the first scale encoding module to obtain the dependency relationship in the time dimension; and generate the first scale latent variable by modeling the obtained dimensional collaborative relationship and time dependency relationship and gradually denoising through stream matching sampling.

[0162] In other embodiments, the above processing module is specifically used to: perform feature transformation on the first scale latent variable through the condition extraction network of the second scale encoding module to generate a trend condition vector; concatenate the trend condition vector with the text condition vector to obtain an enhanced condition vector; and perform multi-step iterative denoising processing based on flow matching on the second random noise based on the enhanced condition vector to generate a second scale latent variable corresponding to the first scale latent variable.

[0163] In other embodiments, the textual sequential diffusion model further includes: a word segmentation module; the aforementioned processing module is specifically used to: perform word segmentation processing on the text prompt words through the word segmentation module to obtain the text word segmentation sequence; input the text word segmentation sequence into the text encoding module, and extract text semantic features through the multi-layer self-attention mechanism of the text encoding module to obtain the text conditional vector.

[0164] In other embodiments, the above-mentioned processing module is specifically used for: performing wavelet decomposition on the time-series training data to obtain low-frequency trend components and high-frequency detail components; processing the low-frequency trend components through the first scale encoding module of the dual-latent spatial variational autoencoder to obtain first-scale latent variable distribution parameters, and processing the low-frequency trend components through the second scale encoding module of the dual-latent spatial variational autoencoder to obtain second-scale latent variable distribution parameters; obtaining first-scale training latent variables and second-scale training latent variables based on the first-scale latent variable distribution parameters and the second-scale latent variable distribution parameters, respectively; performing decoding and reconstruction processing on the first-scale training latent variables and the second-scale training latent variables through the scale decoding module of the dual-latent spatial variational autoencoder to obtain low-frequency trend components and high-frequency trend components, and calculating reconstruction loss; and updating the parameters of the dual-latent spatial variational autoencoder based on the reconstruction loss.

[0165] In other embodiments, the above processing module is further configured to: freeze the updated parameters of the dual latent spatial variational autoencoder; construct multidimensional temporal latent variable training samples based on the first-scale training latent variables and the second-scale training latent variables; train the first-scale encoding module and the second-scale encoding module based on the multidimensional temporal latent variable training samples, and update the parameters of the first-scale encoding module and the second-scale encoding module respectively using a flow matching loss function.

[0166] It should be noted that the time-series data generation device can implement all the processes implemented in the above method embodiments and achieve the same beneficial effects. To avoid repetition, it will not be described again here.

[0167] In the case where the functions of the integrated modules described above are implemented in hardware, this application provides a possible structural schematic diagram of the electronic device involved in the above embodiments. For example... Figure 15 As shown, the electronic device 90 includes: a processor 92, a communication interface 93, and a bus 94. Optionally, the electronic device 90 may also include a memory 91.

[0168] Processor 92 may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 92 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 92 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0169] Communication interface 93 is used to connect with other devices via a communication network. This communication network can be Ethernet, wireless access network, wireless local area network (WLAN), etc.

[0170] The memory 91 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0171] As one possible implementation, the memory 91 can exist independently of the processor 92. The memory 91 can be connected to the processor 92 via a bus 94 and is used to store instructions or program code. When the processor 92 calls and executes the instructions or program code stored in the memory 91, it can implement the timing data generation method provided in the embodiments of this application.

[0172] In another possible implementation, memory 91 can also be integrated with processor 92.

[0173] Bus 94 can be an Extended Industry Standard Architecture (EISA) bus, etc. Bus 94 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 15 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0174] Through the above description of the implementation methods, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the service calling device can be divided into different functional modules to complete all or part of the functions described above.

[0175] This application also provides a computer-readable storage medium. All or part of the processes in the above method embodiments can be executed by computer instructions instructing related hardware. The program can be stored in the aforementioned computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. The computer-readable storage medium can be any of the foregoing embodiments or memory. The aforementioned computer-readable storage medium can also be an external storage device of the aforementioned service invocation device, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the aforementioned service invocation device. Further, the aforementioned computer-readable storage medium can include both internal storage units of the aforementioned service invocation device and external storage devices. The aforementioned computer-readable storage medium is used to store the aforementioned computer program and other programs and data required by the aforementioned service invocation device. The aforementioned computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0176] This application also provides a computer program product, which includes a computer program that, when run on a computer, causes the computer to execute any of the timing data generation methods provided in the above embodiments.

[0177] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating time-series data, characterized in that, include: The text prompt words are input into the text generation time-series diffusion model, which includes at least a dual latent space variational autoencoder, which includes a text encoding module, a scale encoding module, a noise module, and a scale decoding module. The text encoding module encodes the text segmentation sequence corresponding to the text prompt word to obtain a text condition vector; The text conditional vector and the first random noise are sampled and processed by the scale encoding module to generate a target scale latent variable, which represents the feature information of the temporal features contained in the text prompt word; The target scale latent variable is denoised by the denoising module of the textual time-series diffusion model, and the denoised target scale latent variable is input into the scale decoding module for decoding and reconstruction to generate multi-dimensional time-series data. The multi-dimensional time-series data includes the statistical and temporal characteristics of the text prompt words, as well as the dependency between the statistical and temporal characteristics.

2. The method according to claim 1, characterized in that, The scaling module includes a first scaling module, a second scaling module, and a fusion module; the step of sampling the text conditional vector and the first random noise through the scaling module to generate target scale latent variables includes: The first scale encoding module performs stream matching sampling on the text condition vector and the first random noise to generate a first scale latent variable, which represents the low-frequency trend features of the temporal features contained in the text prompt word. The second-scale encoding module performs conditional flow matching sampling on the text conditional vector, the first-scale latent variable, and the second random noise to generate a second-scale latent variable. The second-scale latent variable represents the high-frequency detail features of the temporal features contained in the text prompt word. The target scale latent variable is obtained by synthesizing the first-scale latent variable and the second-scale latent variable through the fusion module.

3. The method according to claim 1, characterized in that, The process involves inputting the processed target-scale latent variables into the scale decoding module for decoding and reconstruction to generate multidimensional time-series data, including: The target scale latent variable is decoded by the scale decoding module to reconstruct low-frequency trend components and high-frequency detail components; wherein, the low-frequency trend components contain trend feature information of the temporal features of the text prompt words, and the high-frequency detail components contain detail feature information of the temporal features of the text prompt words. The low-frequency trend component and the high-frequency detail component are combined to obtain the multidimensional time series data.

4. The method according to claim 2, characterized in that, The step of performing stream matching sampling on the text condition vector and the first random noise through the first scale encoding module to generate the first scale latent variable includes: The text conditional vector is combined with the first random noise by the first scale encoding module to obtain the initial latent variable sequence; The initial latent variable sequence is segmented along its dimensional direction by the first scale encoding module to obtain a dimensional sequence, and the dimensional sequence is rearranged into a time window sequence. The dimensional sequence is processed by the first diffusion transformer of the first scale encoding module to obtain the collaborative relationship between different dimensions; The time window sequence is processed by the second diffusion transformer of the first scale encoding module to obtain the time-dimensional dependencies; Based on the acquired dimensional collaborative relationship modeling and time dependency relationship, noise is gradually removed through flow matching sampling to generate the first scale latent variable.

5. The method according to claim 2, characterized in that, The step of performing conditional stream matching sampling on the text conditional vector, the first-scale latent variable, and the second random noise through the second-scale encoding module to generate the second-scale latent variable includes: The first-scale latent variable is transformed by the conditional extraction network of the second-scale encoding module to generate a trend conditional vector. The trend condition vector is concatenated with the text condition vector to obtain the enhanced condition vector; Based on the enhanced conditional vector, a multi-step iterative denoising process based on flow matching is performed on the second random noise to generate a second-scale latent variable corresponding to the first-scale latent variable.

6. The method according to claim 1, characterized in that, The text-based temporal diffusion model further includes: a word segmentation module; the text encoding module encodes the text segmentation sequence corresponding to the text prompt word to obtain a text condition vector, including: The text prompt words are segmented using the word segmentation module to obtain the text segmentation sequence; The text segmentation sequence is input into the text encoding module, and the text semantic features are extracted through the multi-layer self-attention mechanism of the text encoding module to obtain the text condition vector.

7. The method according to any one of claims 1 to 6, characterized in that, The method further includes: Wavelet decomposition is performed on the time-series training data to obtain low-frequency trend components and high-frequency detail components; The low-frequency trend component is processed by the first-scale encoding module of the dual latent spatial variational autoencoder to obtain the first-scale latent variable distribution parameters, and the low-frequency trend component is processed by the second-scale encoding module of the dual latent spatial variational autoencoder to obtain the second-scale latent variable distribution parameters. The first-scale training latent variables and the second-scale training latent variables are obtained based on the distribution parameters of the first-scale latent variables and the distribution parameters of the second-scale latent variables, respectively. The first-scale training latent variable and the second-scale training latent variable are decoded and reconstructed by the scale decoding module of the dual latent spatial variational autoencoder to obtain low-frequency trend components and high-frequency trend components, and the reconstruction loss is calculated. Based on the reconstruction loss, the parameters of the dual latent space variational autoencoder are updated.

8. The method according to claim 7, characterized in that, After updating the parameters of the dual-latency spatial variational autoencoder based on the reconstruction loss, the method further includes: Freeze the parameters of the updated dual latent space variational autoencoder; Based on the first-scale training latent variables and the second-scale training latent variables, a multidimensional time-series latent variable training sample is constructed. Based on the multidimensional temporal latent variable training samples, the first scale encoding module and the second scale encoding module are trained, and the parameters of the first scale encoding module and the second scale encoding module are updated respectively using the flow matching loss function.

9. A time-series data generation device, characterized in that, include: The processing module is used for: The text prompt words are input into the text generation time-series diffusion model, which includes at least a dual latent space variational autoencoder, which includes a text encoding module, a scale encoding module, a noise module, and a scale decoding module. The text encoding module encodes the text segmentation sequence corresponding to the text prompt word to obtain a text condition vector; The text conditional vector and the first random noise are sampled and processed by the scale encoding module to generate a target scale latent variable, which represents the feature information of the temporal features contained in the text prompt word; The target scale latent variable is denoised by the denoising module of the textual time-series diffusion model, and the denoised target scale latent variable is input into the scale decoding module for decoding and reconstruction to generate multi-dimensional time-series data. The multi-dimensional time-series data includes the statistical and temporal characteristics of the text prompt words, as well as the dependency between the statistical and temporal characteristics.

10. An electronic device, characterized in that, The device includes a processor and a memory, the processor being coupled to the memory; the memory is used to store computer instructions, which are loaded and executed by the processor to enable the computer device to implement the timing data generation method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes computer-executable instructions that, when executed on a computer, cause the computer to perform the time-series data generation method as described in any one of claims 1 to 8.

12. A computer program product, characterized in that, The computer program product includes a computer program that, when run on an electronic device, causes the electronic device to perform the timing data generation method as described in any one of claims 1 to 8.

Images