Interpretable text semantic driving time sequence generation method based on Diffusion Transform model

Abstract

The invention discloses an interpretable text semantic driving time sequence generation method based on a Diffusion Transform model, and relates to the technical field of artificial intelligence and time sequence generation methods. The method comprises the following steps: converting a natural language condition into a control signal generated for a time sequence; dividing a time sequence characteristic explicit into three parts, namely a trend part, a season part and a residual error part, and capturing a long-term time dependence and periodic mode in combination with a time sequence Transform; a time sequence conforming to text semantics is obtained through an iterative sampling process of DDPM, and interpretable trend, season and residual component decomposition results of the time sequence are output at the same time; text condition coding and time embedding are fused into a joint condition vector, intermediate features of a denoising network are modulated through self-adaptive normalization, so that the generation process is dynamically guided by text semantics, and an interpretable text time sequence conforming to an instruction is generated. According to the method, the stability and diversity of the generation process can be improved, and the interpretability and analyzability of the generation result are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and time series generation methods, and in particular to an interpretable text semantic-driven time series generation method based on the Diffusion Transformer model. Background Technology

[0002] Existing methods for time series data generation can be categorized into three types: methods based on Variational Autoencoders (VAEs), methods based on Generative Adversarial Networks (GANs), and methods based on Diffusion Models. VAE-based methods typically use learned approximate inference to efficiently generate synthetic samples; the inference problem uses the values ​​of certain variables or probability distributions to predict other values ​​or probability distributions. GAN-based methods achieve implicit modeling of complex distributions through adversarial training between the generator and discriminator. Time series GANs can be broadly classified into discrete (suitable for discrete time-point data such as event sequences and transaction records) and continuous (suitable for continuous time series such as sensor signals and meteorological data). Diffusion Models-based methods reconstruct high-quality samples from random noise through a progressive noise addition and reverse denoising process. Existing text-based time series generation methods based on diffusion models can also be categorized into three types: label-based conditions, text-based conditions, and researchers are attempting to introduce high-level semantic information into the time series generation process and cross-domain text-guided time series generation.

[0003] Variational autoencoder-based methods tend to oversmooth when generating time series with complex long-range dependencies or strong noise perturbations. Generative adversarial network-based methods suffer from unstable training processes, prone to pattern collapse leading to a lack of diversity in generated sequences. Diffusion model-based methods suffer from low generation efficiency, high inference costs, difficulty in direct semantic-level control, and limited interpretability of modeling real-world dynamics in complex scenarios, failing to meet the stringent requirements of real-time performance and controllability in practical applications. Text-based time series generation methods also have shortcomings: label-conditional methods lack deep semantic understanding; text-conditional methods are only applicable to specific domains; and existing cross-domain text-guided time series generation methods suffer from a lack of interpretable generation processes and insufficient generation accuracy. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to provide an interpretable text temporal generation method that can enhance long-term dependencies, improve the stability and diversity of the generation process, achieve precise control of temporal generation by cross-domain text semantics, and improve the interpretability and analyzability of the generation results.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an interpretable text semantic-driven temporal generation method based on the DiffusionTransformer model, comprising the following steps: S1, Text Temporal Alignment: The input natural language text conditions are transformed into control signals for the temporal generation process. The text semantics are globally modulated on the feature channels through channel gating, and the local dynamic alignment of the text description with the specific moment of the temporal sequence is achieved through a cross-modal attention mechanism. S2, Interpretable Time Series Decomposition: A trend modeling module and a seasonality modeling module are set up in parallel within the denoising network. In each step of the denoising process, the trend component and seasonal component of the generated sequence are extracted to achieve interpretable decomposition. S3, Text Conditional Diffusion Training and Interpretable Output: Using DDPM as a framework, the model directly predicts clean samples, performs noise calculation, and obtains time series that conform to text semantics through the iterative sampling process of DDPM, while simultaneously outputting its interpretable trend, seasonal and residual component decomposition results. S4, Modulation Diffusion: The text conditional encoding and temporal embedding are fused into a joint conditional vector. The intermediate features of the modulation denoising network are adaptively normalized to guide the generation process dynamically by the text semantics, generating interpretable text temporal sequences that conform to the instructions.

[0006] The beneficial effects of adopting the above technical solution are as follows: 1) High-fidelity generation and high semantic consistency: Through text-dimensional gating modulation and cross-modal cross-attention structure, a multi-level accurate mapping from text semantics → channel modulation → local alignment is achieved. Finally, through the iterative sampling process of DDPM, a high-quality time series that conforms to the text semantics is obtained.

[0007] 2) Component-level structure interpretability: A trend-seasonal decomposition mechanism is explicitly introduced into the time-series Transformer architecture, decomposing the time series into three parts: trend, season, and residuals. The model extracts and reconstructs these components in parallel through multiple blocks, combining time-domain and frequency-domain joint losses. This achieves component-level interpretability and effectively avoids the defect of traditional diffusion models that destroy periodic information due to noise.

[0008] 3) Overcoming the bottleneck of long-term dependency modeling: The U-Net backbone of the traditional diffusion model is replaced by a temporal Transformer, which fully utilizes the self-attention mechanism of the Transformer architecture in sequence modeling. This effectively overcomes the bottleneck of long-term dependency modeling of traditional one-dimensional convolution or RNN, and can capture the complex dynamic relationships and long-term statistical patterns across arbitrary distances in time-series data.

[0009] 4) Precise and controllable generation process: By introducing the Classifier-Free Guidance (CFG) mechanism and using the difference between conditional and unconditional predictions for amplification and guidance during the sampling phase, the control effect of text semantics is significantly enhanced, making the generation process more precise and controllable.

[0010] 5) Joint modeling and flexible modulation: The adaptive layer normalization (AdaLN) mechanism is adopted, which enables the model to flexibly adjust the distribution of generated features according to the time step and text semantics during the diffusion process, realizing joint modeling of text semantics and temporal dynamics. Attached Figure Description

[0011] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0012] Figure 1 This is the main flowchart of the method described in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the system described in Embodiment 2 of the present invention; Figure 3 This is a detailed structural diagram of the system described in Embodiment 2 of the present invention; Figure 4 Flowchart of the text conditional and time step fusion modulation mechanism in the diffusion model; Figure 5 A diagram of trend modeling modules in an interpretable temporal decomposition structure; Figure 6 This is a diagram of seasonal modeling modules in an interpretable temporal decomposition structure. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0014] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0015] Example 1: Overall, such as Figure 1This invention discloses an interpretable text semantic-driven temporal generation method based on the Diffusion Transformer model, the method comprising the following steps: S1, Text Temporal Alignment: The input natural language text conditions are transformed into control signals for the temporal generation process. The text semantics are globally modulated on the feature channels through channel gating, and the local dynamic alignment of the text description with the specific moment of the temporal sequence is achieved through a cross-modal attention mechanism. S2, Interpretable Time Series Decomposition: A trend modeling module and a seasonality modeling module are set up in parallel within the denoising network. In each step of the denoising process, the trend component and seasonal component of the generated sequence are extracted to achieve interpretable decomposition. S3, Text Conditional Diffusion Training and Interpretable Output: Using DDPM as a framework, the model directly predicts clean samples, performs noise calculation, and obtains time series that conform to text semantics through the iterative sampling process of DDPM, while simultaneously outputting its interpretable trend, seasonal and residual component decomposition results. S4, Modulation Diffusion: The text conditional encoding and temporal embedding are fused into a joint conditional vector. The intermediate features of the modulation denoising network are adaptively normalized to guide the generation process dynamically by the text semantics, generating interpretable text temporal sequences that conform to the instructions.

[0016] The proposed method can enhance long-term dependencies and improve the stability and diversity of the generation process; it can achieve precise control of time-series generation through cross-domain text semantics, maintain high accuracy while clearly decomposing factors such as trends, seasonality and random disturbances, and improve the interpretability and analyzability of the generation results.

[0017] Example 2: like Figures 2-3 As shown, this embodiment of the invention discloses an interpretable text temporal generation system based on a diffusion model, comprising: Text temporal alignment module 101: This module is responsible for converting the input natural language text conditions into signals used to control the temporal generation process. It achieves global modulation of feature channels by text semantics through a channel gating mechanism and utilizes a cross-modal attention mechanism to achieve local dynamic alignment between the text description and specific moments in the temporal sequence. Interpretable temporal decomposition module 102: This module is based on the temporal Transformer architecture. During the diffusion denoising process, the temporal features are explicitly decomposed into trend components, seasonal components and residual components. The trend components are modeled by a multinomial regressor, the seasonal components are reconstructed by dynamically selecting the main frequency components after Fourier transform, and the residual components are processed by a self-attention mechanism. The model uses the sum of the three as the prediction of clean samples and adopts the training process constrained by the joint loss of the time domain and frequency domain. Text Conditional Diffusion Training and Interpretable Output Module 103: This module fuses temporal embedding and text conditional encoding at each diffusion time step to generate a joint conditional vector, which is then integrated into each time step of the denoising network through an adaptive layer normalization mechanism to achieve dynamic modulation of the feature distribution. In the sampling stage, a classifier-free guidance technique is used to enhance the text control strength, and finally, a high-quality time series that conforms to the text semantics is output.

[0018] Figure 4 The text conditional and temporal step fusion modulation mechanism in the diffusion model is illustrated. The temporal step embedding and text vector are first fused together into a joint conditional vector, which is then passed through a multilayer perceptron to output multiple sets of scaling parameters, which are used to modulate multiple normalization layers within the denoising network.

[0019] Figure 5 The trend modeling module in an interpretable time series decomposition structure is illustrated. This module is designed to process features within the denoising network through convolutional layers, followed by regression modeling using a multinomial basis. Features first pass through the first convolutional layer, then undergo transposition and pass through an activation function, followed by the second convolutional layer, and finally perform a specific operation (represented by the "@" symbol) with the multinomial basis to construct a smooth trend component that captures low-frequency variations in the time series.

[0020] Figure 6 This paper illustrates a seasonality modeling module within an interpretable time-series decomposition structure. This module explicitly extracts periodic components using Fourier analysis. First, a Fourier transform is performed on the input time-series features to obtain frequency components, which are categorized into odd-length and even-length frequency components based on their length. Subsequently, the module selects the Top-K frequency components with the largest amplitudes and their conjugate pairs, then performs frequency amplification on the selected frequencies. Finally, an inverse Fourier transform is used to reconstruct the seasonality components with periodic waveforms.

[0021] The above steps will be explained in detail below with specific methods: Text temporal alignment module 101: Effectively converts natural language conditions into control signals for temporal generation, possessing both global / hierarchical modulation and time-by-time alignment capabilities. In the gating mechanism, text vectors... First, it undergoes linear projection. A representation consistent with the feature channel dimension is obtained, and then compared with the learnable channel-gated vectors. Multiplication generates the final gating signal Meanwhile, the model introduces a cross-modal attention structure in each layer: time series features are used as queries, text embeddings are used as keys / values, and attention weights are used to determine which descriptions in the text are relevant to which moments in the time series.

[0022] Interpretable Temporal Decomposition Module 102: Explicitly divides temporal features into three parts: trend, seasonality, and residuals, and combines a temporal transformer to capture long-term time dependencies and cyclical patterns. The design includes: First, replacing the U-Net backbone of the diffusion model with a temporal transformer, making the denoising network a stack of pure Transformer modules; designing a three-branch parallel decomposition structure, from an interpretability perspective, for time series... Approximate decomposition into: ; in For trends, For seasonality, For residuals, these components are actively separated from noise and reconstructed by forcing the model. The architecture of the temporal Transformer is fully utilized to model long-term time dependencies, and trends and seasons are explicitly modeled by different sub-modules.

[0023] trend Modeled using a multinomial regressor, by applying the first... layer The channel average value output at time step t Apply a linear mapping and use a polynomial basis with the normalized time vector. Perform element-wise tensor multiplication, where... It is a time axis coordinate. It is the specified polynomial order. This is used to construct smooth low-frequency trend components. ; seasonal By the layer Output at time step t Perform a discrete Fourier transform and dynamically select the one with the largest amplitude. Frequency components and their conjugate pairs, reconstructing periodic waveforms This enables explicit modeling of complex seasonal patterns and incorporates residuals. Retain the self-attention processing. (Trend) Seasonal With residual The original signal is obtained by combining the three components. The formula is as follows: ; in, Represents the original sequence go through The sequence obtained after forward noise addition. This represents the set of learnable parameters for the denoising network.

[0024] By extracting component representations from multiple blocks in parallel, structured dynamics are actively reconstructed from noise, and joint expectation is used. Calculate time-frequency domain loss The loss formula is as follows: .

[0025] in These are the weights used to balance the "time-domain reconstruction error" and the "frequency-domain reconstruction error". These are time step weighting coefficients. This represents the Fast Fourier Transform.

[0026] Text Conditional Diffusion Training and Interpretable Output Module 103: This invention uses DDPM as a framework, where the model directly predicts clean samples and then calculates noise. In the diffusion process... Step, for the original time series Forward noise addition yields a noisy sequence ,in , Standard Gaussian noise, , They were respectively through The signal-to-noise ratio parameter accumulated step by step. Time step and text condition vector Input a time-series Transformer backbone model, and the model predicts the original signal. We obtain the result by summing the three components: ; in, The modeling processes for the trends, seasons, and residuals mentioned above are respectively. Based on the predicted... Calculate the corresponding noise And calculate noise error : ; During the sampling phase, noise is generated through iterative DDPM denoising. : ; Introducing Classifier-Free Guidance to guide noise The calculation is as follows: ; in For guiding strength coefficient, and Using and not using text conditions respectively The noise in the time model prediction is eliminated. Finally, through the iterative sampling process of DDPM, a high-quality time series that conforms to the semantics of the text is obtained, and its interpretable trend, seasonal and residual component decomposition results are output simultaneously.

[0027] The text diffusion process is modulated through adaptive layer normalization: the model modulates the diffusion process at each diffusion time step. Embedding time Combined with text conditional encoding to generate joint conditional vectors This is used to guide the feature modulation of the denoising network. Specifically, through the adaptive AdaLN mechanism, the joint condition vector is... Projection is the scaling parameter of the feature channel. With scaling parameters ,Right now The normalization results of each layer of the Transformer are conditionally adjusted as follows: ; in and They are respectively The mean and standard deviation along the feature dimension cause the feature distribution of each denoising step to be modulated by both text and time.

[0028] Application Example 1: The method provided by this invention uses a high-fidelity sea clutter simulator as an example to generate I / Q sampling time series with specific environmental characteristics based on environmental description text. The deployment of the model enables it to generate sea clutter signals conforming to specific semantics based on sea state text conditions (such as wind speed, wave level, periodic trends, etc.).

[0029] The model is deployed on a high-performance computing cluster and serves as the core signal generation engine. The backbone of the denoising network is an interpretable temporal Transformer, which replaces the U-Net backbone of traditional diffusion models. The model receives two types of input: encoded sea state text conditions. and noisy time series samples The core of this generation method lies in the sea state text conditions. The introduction and modulation of [the technology], through an adaptive layer normalization mechanism, enables precise control of the temporal generation of text.

[0030] This model acts as a signal generation engine, working in conjunction with an external text encoding module and a radar simulation platform. It can quickly generate multiple sets of high-quality sea clutter sequences that conform to textual semantics and simultaneously output their trends. ,season and residual The decomposition results are provided for radar engineers to perform component-level analysis.

[0031] Implementation steps: (1) Model installation and training preparation: First, the key structures of this invention need to be deployed, including text-dimensional gating modulation for text guidance. The system includes a cross-modal attention structure for fine-grained alignment, a three-branch parallel decomposition structure for interpretability, and an adaptive layer normalization mechanism for feature conditional adjustment. After component deployment, the forward noise addition process under the DDPM framework is configured. During the training phase, the model is endowed with unconditional generation capabilities by randomly masking text conditions. This is in preparation for introducing Classifier-Free Guidance (CFG) guidance during the sampling phase.

[0032] (2) Core component debugging and optimization: During the debugging process, the core task is to achieve accurate alignment and decomposition of text and time-series data. The first step is to debug the gating signal. And cross-attention weights are used to ensure that text semantics can effectively modulate feature channels and achieve fine-grained semantic alignment with time series. Secondly, the decomposition structure needs to be tuned, including the multinomial regressor used for trend modeling. Discrete Fourier Transform for Seasonal Modeling and Selection of its Top-K Frequency Components This ensures that the model can actively separate and reconstruct these structured dynamics from noise. Finally, a joint time-domain and frequency-domain loss is employed. Training is performed to ensure that the sequences generated by the model are It combines high-fidelity generation with component-level interpretability.

[0033] (3) Model operation and maintenance: During the sampling phase, the model receives noisy samples. and text conditions Through the AdaLN mechanism After feature modulation, the guiding noise is calculated iteratively. Sampling was performed to obtain a time series that conformed to the text semantics. The model also outputs its interpretable trend, seasonal, and residual decomposition results. In terms of maintenance, the accuracy of the decomposition components needs to be checked periodically, and the model needs to be fine-tuned based on new sea state data distributions to continuously maintain the model's generation quality and semantic consistency.

[0034] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0035] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units, etc., and are not limited to these.

[0036] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0037] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for interpretable text semantic-driven temporal generation based on the Diffusion Transformer model, characterized in that... Includes the following steps: S1, Text Temporal Alignment: The input natural language text conditions are transformed into control signals for the temporal generation process. The text semantics are globally modulated on the feature channels through channel gating, and the local dynamic alignment of the text description with the specific moment of the temporal sequence is achieved through a cross-modal attention mechanism. S2, Interpretable Time Series Decomposition: A trend modeling module and a seasonality modeling module are set up in parallel within the denoising network. In each step of the denoising process, the trend component and seasonal component of the generated sequence are extracted to achieve interpretable decomposition. S3, Text Conditional Diffusion Training and Interpretable Output: Using DDPM as a framework, the model directly predicts clean samples, performs noise calculation, and obtains time series that conform to text semantics through the iterative sampling process of DDPM, while simultaneously outputting its interpretable trend, seasonal and residual component decomposition results. S4, Modulation Diffusion: The text conditional encoding and temporal embedding are fused into a joint conditional vector. The intermediate features of the modulation denoising network are adaptively normalized to guide the generation process dynamically by the text semantics, generating interpretable text temporal sequences that conform to the instructions.

2. The interpretable text semantic-driven temporal generation method based on the Diffusion Transformer model as described in claim 1, characterized in that, The text timing alignment includes the following steps: Text vector First, it undergoes linear projection. A representation consistent with the feature channel dimension is obtained, and then compared with the learnable channel-gated vectors. Multiplication generates the final gating signal ; Time series features are used as queries, and text embeddings are used as keys / values. Attention weights are used to determine which descriptions in the text are relevant to which moments in the time series.

3. The interpretable text semantic-driven temporal generation method based on the Diffusion Transformer model as described in claim 1, characterized in that, Interpretable time series decomposition methods include the following steps: First, the U-Net backbone of the diffusion model is replaced with a temporal Transformer, making the denoising network a stack of pure Transformer modules; a three-branch parallel decomposition structure is designed for time series... Approximate decomposition into: ； in For trends, For seasonality, For the residuals, these components are actively separated from the noise and reconstructed by forcing the model; trend Modeled using a multinomial regressor, by applying the first... layer The channel average value output at time step t Apply a linear mapping and use a polynomial basis with the normalized time vector. Perform element-wise tensor multiplication, where... It is a time axis coordinate. It is a specified polynomial order, used to construct smooth low-frequency trend components. ; seasonal By the layer Output at time step t Perform a discrete Fourier transform and dynamically select the one with the largest amplitude. Frequency components and their conjugate pairs, reconstructing periodic waveforms This enables explicit modeling of complex seasonal patterns and incorporates residuals. Retain the self-attention processing, and the trend Seasonal With residual The original signal is obtained by combining the three components. The formula is as follows: ； in, Represents the original sequence go through The sequence obtained after forward noise addition. This represents the set of learnable parameters for the denoising network; By extracting component representations from multiple blocks in parallel, structured dynamics are actively reconstructed from noise, and joint expectation is used. Calculate time-frequency domain loss The loss formula is as follows: ； in These are the weights used to balance the time-domain reconstruction error and the frequency-domain reconstruction error. These are time step weighting coefficients. This represents the Fast Fourier Transform.

4. The interpretable text semantic-driven temporal generation method based on the Diffusion Transformer model as described in claim 3, characterized in that, The method for text conditional diffusion training and interpretable output includes the following steps: In the diffusion process Step, for the original time series Forward noise addition yields a noisy sequence ,in , Standard Gaussian noise, , They were respectively through The cumulative signal-to-noise ratio parameter; Time step and text condition vector Input a time-series Transformer backbone model, and the model predicts the original signal. We obtain the result by summing the three components: ； in, The modeling processes for the trends, seasons, and residuals mentioned above are respectively; based on the predictions... Calculate the corresponding noise And calculate noise error : ； During the sampling phase, noise is generated through iterative DDPM denoising. : ； Introducing Classifier-Free Guidance to guide noise The calculation is as follows: ； in For guiding strength coefficient, and Using and not using text conditions respectively Noise in the model predictions; Finally, through the iterative sampling process of DDPM, a high-quality time series that conforms to the semantics of the text is obtained, and its interpretable trend, seasonal and residual component decomposition results are output simultaneously.

5. The interpretable text semantic-driven temporal generation method based on the Diffusion Transformer model as described in claim 1, characterized in that, The modulation diffusion method includes the following steps: By using the adaptive AdaLN mechanism, the joint condition vector Projection is the scaling parameter of the feature channel. With scaling parameters ,Right now ; The normalization results of each layer of the Transformer are conditionally adjusted as follows: ； in and They are respectively The mean and standard deviation along the feature dimension cause the feature distribution of each denoising step to be modulated by both text and time.

Images