DAS noise modeling method based on conditional diffusion model

Abstract

The invention relates to a DAS noise modeling method based on a conditional diffusion model, and belongs to the technical field of distributed optical fiber sensing data processing. DAS noise data is acquired and analyzed, a DAS-conditional diffusion model is constructed, and model training and data generation are carried out. According to the method, the statistical characteristics and conditional constraints of the DAS noise are fused in the diffusion process, so that the physical consistency and diversity of the generated noise are effectively improved, and the problem of lack of high-quality mark data in the DAS system is solved. Experiments show that the generated noise and the real noise are highly similar in time-frequency domain and statistical characteristics, the performance of the denoising network based on the extended training set is remarkably improved, high-quality data support is provided for tasks such as DAS data denoising and signal enhancement, and the application capability of the DAS technology in a complex environment is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of Distributed Acoustic Sensing (DAS) data processing technology, and particularly relates to a DAS noise modeling method based on a conditional diffusion model, which is applicable to tasks such as noise data generation, training set expansion, and subsequent signal denoising in DAS systems. Background Technology

[0002] In DAS systems, noise has always been a key factor affecting data quality and signal processing efficiency. In recent years, deep learning-based methods have shown great potential in DAS data denoising. Due to their powerful representational capabilities and training on large datasets driven by numerous learnable parameters, learning-based methods can adaptively learn and capture complex patterns and features in the data, thus achieving more efficient and accurate denoising. However, the success of these methods heavily relies on high-quality training datasets. Low-quality datasets (data with limited noise types, inaccurate labeling, or containing artifacts) can cause biases in the network's learning process, leading to decreased denoising performance. While removing noise, useful signal information may be lost, or artifacts and distortions may be introduced into the denoising results. Furthermore, insufficient training data limits the model's adaptability to different physical conditions, causing the network to learn only partial noise features. This results in poor generalization ability of the denoising network when facing diverse and complex noise environments in DAS data, increasing the model's dependence on specific work area scenarios and weakening its robustness and portability in new environments.

[0003] However, collecting and labeling large amounts of high-quality DAS data is currently time-consuming and expensive, and the data quality is difficult to guarantee. Training set creation methods relying on real-world data collection and manual annotation struggle to cover complex noise scenarios, resulting in insufficient generalization ability in denoising tasks. Especially in DAS systems, environmental noise, instrument noise, and fiber optic loss all contribute to increased noise complexity, further limiting the effectiveness of traditional datasets. Furthermore, the high cost and limited coverage of DAS data collection mean that existing methods cannot meet the diverse data requirements of deep learning models. Therefore, establishing synthetic datasets that reflect different noise characteristics and complex environments is crucial. By utilizing noise modeling techniques to generate diverse noise data, we can not only enrich the training set and reduce data collection costs but also significantly improve the model's generalization performance, thus providing more comprehensive data support for tasks such as DAS denoising and signal enhancement.

[0004] Currently, methods for modeling DAS noise mainly fall into two categories: numerical simulation modeling methods and data-driven generative modeling methods. Numerical simulation methods utilize numerical techniques such as the finite difference method and the finite element method to construct complex DAS noise scenarios based on physical models. Numerical simulation methods can generate realistic noise fields by simulating wave propagation, refraction, and reflection. Their advantage lies in their ability to capture the essential characteristics of the physical system relatively well, and the generated noise exhibits a certain degree of physical consistency. However, these methods often rely on physical assumptions and struggle to fully reflect the diversity and complexity of DAS noise. Furthermore, their computational complexity is often very high, requiring substantial computational resources, and the effectiveness of the models constructed decreases significantly when faced with complex initial and boundary conditions.

[0005] Generative models based on Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) have made significant progress in noise modeling in recent years. These data-driven methods learn the noise distribution of real-world data to generate similar noise data. These methods can be trained on large amounts of real-world data, automatically learning and capturing complex noise distribution features without relying on explicit physical models. They generate realistic data while exhibiting good scalability and adaptability, suitable for modeling various noise types. However, GANs and similar models may suffer from mode collapse or loss of detail when generating DAS noise, resulting in limited noise variety and quality. VAEs typically use a simple Gaussian distribution to approximate the real noise distribution. This overly simplistic distribution assumption fails to capture the complex features of DAS noise, causing the generated noise to not perfectly match the actual noise situation. These methods often struggle to accurately capture the complex spatial correlations and multi-scale features in DAS noise. Summary of the Invention

[0006] This invention provides a DAS noise modeling method based on a conditional diffusion model to address the problems of low generated noise fidelity, poor physical consistency, difficulty in covering complex noise types, and the performance limitations of deep learning models due to the scarcity of DAS training data in existing DAS noise modeling methods. By constructing a DAS noise modeling method based on a conditional diffusion model, incorporating the statistical characteristics and conditional constraints of DAS noise, a diverse range of noise data highly similar to real noise is generated, expanding the training set, improving the performance of subsequent denoising networks, and providing an efficient solution for DAS data processing.

[0007] The technical solution adopted by this invention includes the following steps:

[0008] (1) Acquire and analyze DAS noise data;

[0009] (2) Constructing the DAS-conditional diffusion model;

[0010] (3) Model training and data generation.

[0011] Step (1) of the present invention includes:

[0012] 1) Collect DAS noise data;

[0013] Noisy data was collected from the actual DAS system. Background noise and passive data before the arrival of the vibration were selected, and pure noise samples containing only various typical mixed noises and no seismic signal components were extracted to avoid the seismic signal components from interfering with the subsequent analysis.

[0014] 2) Extract four typical noise types from pure noise samples;

[0015] Based on the obtained pure noise samples, four typical DAS noises were separated and extracted using a bandpass filtering method. The four types of noise were then normalized and cropped into 256×256 data blocks to obtain four noise training sets.

[0016] 3) Analyze the statistical characteristics of noise;

[0017] Statistical characteristics of the four typical noises were analyzed, and the mean μ and variance σ² of each type of noise were calculated. At the same time, the spectral characteristics of each type of noise were obtained by short-time Fourier transform, which provides a basis for the subsequent design of diffusion model parameters.

[0018] Step (2) of the present invention includes:

[0019] 1) Design the forward process of the DAS-conditional diffusion model;

[0020] The mean and variance of various types of DAS noise are introduced, and parameterized noise that conforms to the statistical characteristics of DAS noise is continuously added to the original data during the forward process. The conditional distribution of noise addition transformation is performed at each time step. for:

[0021] ;

[0022] Where, x t x t-1 Let these represent the state variables at step t and step (t-1) of the diffusion process, respectively. It follows a mean of 0 and a variance of σ. 2 Gaussian distributed random noise, β t This is the noise scheduling parameter at step t, used to control the degree of noise addition at each step, and it is smoothed by a cosine-increasing form:

[0023] ;

[0024] here, It is the minimum value of the noise scheduling parameter. is the maximum value of the noise scheduling parameter, T is the total number of time steps in the diffusion process, and t is the current time step; where 1≤t≤T;

[0025] Each step of the diffusion process starts from... The mean, The process of sampling noise from a Gaussian distribution of variance and adding it to the data from the previous step, based on the above formula, effectively transforms the clean sample data into another type of noise interpolation step by step. All are derived from the output of the previous step. The result is obtained by weighted summation with noise, from to any time step The conditional distribution is:

[0026] ;

[0027] in, This is the initial sample for the diffusion process; , , representing the cumulative coefficient at different time steps during the diffusion process, used to attenuate the original sample. The influence of DAS-conditional diffusion model The noise is generated based on the statistical characteristics of DAS data, and after a sufficient time step, The process eventually converges to the prior distribution, yielding the final noisy state of the forward process. ;

[0028] 2) Design the reverse process of the DAS-conditional diffusion model;

[0029] Defined on a Markov chain, which undoes the forward diffusion process and performs iterative denoising, its goal is to denoise from the final noisy state. Data distribution recovered , This refers to the generated data obtained from the final modeling of the diffusion model. In the reverse process, the spectral characteristics of the noise samples obtained through short-time Fourier transform are used as condition y to guide the generation of the target data. Under the guidance of y, the reverse process obtains a more accurate target distribution. ;

[0030] A neural network is used to fit the back transition probability. :

[0031] ;

[0032] in, The variance function is fixed, and its value is... , The mean function of the network's learning predictions.

[0033] ;

[0034] Here This refers to the standardized residuals predicted by the neural network. The network is trained to predict noise components at specific time steps. To effectively generate DAS noise data that conforms to physical laws, an improved U-Net network is used. The U-Net network achieves multi-scale feature fusion and reconstruction through symmetrically arranged residual blocks combined with bottom-up skip connections. Finally, the decoder output is processed by the output module to generate the noise components predicted by the network. , to be used in the reverse reconstruction process of the diffusion model;

[0035] During the generation phase, the diffusion model uses an improved U-Net network to progressively eliminate noise in the data, generating data step by step. The optimization function of this network is as follows:

[0036] ;

[0037] After training, the network will master the Gaussian transition probabilities during the reverse process. ,from The generated data is obtained by sampling. :

[0038] ;

[0039] in, The preset reverse process noise scale, It follows a mean of 0 and a variance of . The random variables are Gaussian distributed. Finally, the DAS-conditional diffusion model generates noise-free data step by step through sampling steps. :

[0040] .

[0041] The improved U-Net network in step 2) of this invention includes noisy signal samples at its input. The noise prior information y, which serves as a condition, is initially fused by a 3×3 convolutional layer and then enters multiple residual blocks. The residual blocks extract and map the input features step by step and connect to subsequent modules through feature propagation.

[0042] In the middle section, a temporal embedding module is set up. This temporal embedding module converts the time step t in the diffusion process into the corresponding embedding representation and applies it to the operation of the residual block to achieve adaptation to different diffusion stages. At the same time, an intermediate attention module is set up at the bottleneck position of the network to model global dependencies in the stage where features are most concentrated, thereby improving the ability to capture noise structures. Through this module, the network can identify and utilize long-distance dependencies in DAS data to improve the physical consistency of generated noise.

[0043] The intermediate attention module of this invention generates attention weights by calculating the correlation between the query matrix Q, the key matrix K, and the value matrix V, thereby optimizing feature representation; the calculation process of self-attention is as follows:

[0044] ;

[0045] in, To query the dimensions of the key matrix and the key matrix, Let K be the transpose of the key matrix K.

[0046] Step (3) of the present invention includes:

[0047] 1) Train the DAS-conditional diffusion model;

[0048] Based on the constructed model, the model is trained, and multi-type DAS noise data is generated based on the trained model.

[0049] 2) Generation of DAS noise data;

[0050] A synthetic DAS noise sample consistent with the target category features is obtained.

[0051] The specific training process of step 1) of this invention is as follows:

[0052] First, the four types of DAS noise samples extracted separately and their corresponding noise category labels are used as input data pairs to construct training samples, resulting in four noise training sets. The four noise training sets are divided into training and validation sets in a 7:3 ratio. Each sample is associated with corresponding statistical features, including mean, variance, and spectral feature matrix extracted through short-time Fourier transform.

[0053] For each training step, a time step t is randomly selected, and Gaussian noise related to the time step is added to the original noise sample according to the diffusion process formula, thereby obtaining the noisy sample at that time step. Then, the noisy samples The conditional information y and time step t are input into the U-Net network structure of the conditional diffusion model. The network outputs the corresponding noise prediction result. The loss function is calculated between the prediction result and the known real noise, and the network parameters are updated through backpropagation. Through iterative training with a large number of samples, the model can gradually learn the statistical distribution and category characteristics of DAS noise. The mean absolute error (MAE) is used as the loss function to measure the difference between the generated noise and the real noise.

[0054] .

[0055] Step 2) of the present invention includes generating noise data after the model training is completed. First, random Gaussian noise is used as the initial input, and the conditional information y of the target noise is given. Then, the model iterates from the preset maximum time step T to time step 0 according to the back diffusion sampling process. In each iteration, the model receives the current noisy sample and conditional information, predicts the noise component, and calculates the sample of the previous time step according to the back diffusion formula. After multiple iterations, a synthetic DAS noise sample consistent with the target category features is finally obtained.

[0056] The advantages of this invention are as follows:

[0057] (i) High noise generation accuracy

[0058] This invention incorporates the statistical distribution characteristics of measured DAS noise into the forward diffusion process and constructs a noise sample sequence through gradual perturbation. In the reverse generation process, the improved U-Net network accurately estimates the noise components at each time step. Through large-scale training, the model accurately reproduces the temporal and frequency domain characteristics of typical DAS noise, ensuring that the generated samples are highly consistent with real noise in terms of amplitude probability distribution and spectral energy concentration range. Therefore, this invention has the advantages of high noise generation accuracy and the ability to replace real acquired data.

[0059] (ii) Strong physical consistency

[0060] This invention introduces a diffusion mechanism in its model design that matches the physical process of noise in DAS (Discrete Amplification System). In the forward pass, parameterized noise is injected into the original data, while in the backward pass, the target noise is reconstructed step-by-step by combining time-step embedding and conditional priors. The U-Net network architecture maintains consistency of features at different scales through residual blocks and skip connections, and the intermediate attention module enhances the ability to capture the global noise structure, thus ensuring consistency between the generated results and the actual physical noise in terms of spatiotemporal correlation. This design enables this invention to possess not only numerical accuracy in the noise generation process but also significant advantages in terms of physical interpretability.

[0061] (iii) Excellent adaptability and controllability

[0062] This invention utilizes the spectral characteristics of noise as conditional information input to the model, enabling independent modeling and generation of multiple noise types. Users can select the target noise category as needed, thus achieving controllable generation of multiple noise types within a single model. Furthermore, by adjusting the time step of the diffusion process, noise intensity parameters, and variance perturbation level, the noise intensity and texture details of the generated samples can be flexibly controlled to meet the needs of different experimental or application scenarios.

[0063] (iv) Data augmentation and downstream application effects are significant.

[0064] This invention effectively expands the scale of training data, solving the problems of limited noise types and insufficient sample quantity in traditional DAS datasets. The generated diverse noise data can be used to train denoising models, significantly improving their robustness and signal fidelity in complex noise environments. The noise samples generated by this invention provide more challenging training samples for denoising models, thereby enhancing the model's adaptability to complex noise.

[0065] (v) High value for promotion and application

[0066] This invention has low requirements in terms of data requirements and annotation dependence, making it suitable for applications in scenarios where DAS data is scarce or expensive to acquire. Its method is applicable to noise modeling and signal enhancement in seismic exploration. With its high-precision noise generation, physical consistency, and diverse modeling capabilities, this invention has strong promotion and application value in both academic research and engineering practice. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of four typical noises in the DAS system of this invention. Sub-diagrams (a) to (d) correspond to the original signal forms of random noise, fading noise, horizontal noise, and optical noise, respectively, and intuitively show the characteristic differences of different noises in spatiotemporal distribution.

[0068] Figure 2 This is a schematic diagram of the forward and backward processes of the DAS-conditional diffusion model of the present invention, showing the model from the original noise data. Through the forward noise addition process Noise data is generated step by step. Then, through a reverse denoising process, noise is predicted using the UNET network. ,from Get generated data The complete process;

[0069] Figure 3This is a schematic diagram of the improved U-Net network architecture in the DAS-conditional diffusion model of the present invention, which includes an input module, a residual module, a temporal embedding module, an intermediate self-attention module, a skip connection and an output module. The diagram shows the connection relationship between the network modules.

[0070] Figure 4 This is a schematic diagram of the self-attention module of the present invention, illustrating the calculation and interaction process of the query matrix (Q), key matrix (K), and value matrix (V), which is used to capture long-range dependencies in noisy data;

[0071] Figure 5 The four types of DAS noise generated are shown in sub-figures (a) to (d), which correspond to random noise, fading noise, horizontal noise, and optical noise, respectively.

[0072] Figure 6 The graphs show a comparison of the power spectral density (PSD) of real noise and generated noise. Subgraphs (a) to (d) correspond to random noise, fading noise, horizontal noise, and optical noise, respectively. The horizontal axis represents frequency (Hz), and the vertical axis represents power spectral density (dB). This verifies the advantages of the method of this invention in reproducing the frequency characteristics of noise, especially the significant improvement in the overlap in the mid-to-low frequency band.

[0073] Figure 7 The graphs show the probability distribution of real noise and generated noise. Subgraphs (a) to (d) correspond to random noise, fading noise, horizontal noise, and optical noise, respectively. The horizontal axis represents amplitude (m) and the vertical axis represents probability. The graphs demonstrate the high degree of consistency between the two in terms of distribution trends and high probability regions.

[0074] Figure 8 The images show a comparison of the denoising results of real DAS data. (a) shows the real noisy record; (b) and (c) show the denoising results of the DNCNN network trained on the original small sample training set and the extended training set of this invention, respectively; (d) and (e) show the difference between the corresponding denoising results and the original noisy record, further verifying the ability of the method of this invention to effectively separate signals and noise in real complex noise environments and preserve the key waveform details. Detailed Implementation

[0075] Includes the following steps:

[0076] (1) Acquire and analyze DAS noise data;

[0077] 1) Collect DAS noise data: Collect noisy data from the actual DAS system, select background noise and passive data before the arrival of the vibration, and extract pure noise samples containing only various typical mixed noises and no seismic signal components to avoid the seismic signal components from interfering with subsequent analysis.

[0078] 2) Extracting four typical noise types from pure noise samples: Based on the obtained pure noise samples, four typical DAS noise types are separated and extracted using a bandpass filtering method. Simultaneously, the separated four noise types are normalized and pruned into 256×256 data blocks to obtain four noise training sets, such as... Figure 1 The images show data blocks obtained after trimming four different types of noise.

[0079] 3) Analyze the statistical characteristics of noise. For the four typical noises separated, perform statistical characteristic analysis, calculate the mean μ and variance σ² of each type of noise, and obtain the spectral characteristics of each type of noise through short-time Fourier transform to provide a basis for the subsequent design of diffusion model parameters.

[0080] (2) Constructing the DAS-conditional diffusion model;

[0081] 1) Design the forward process of the DAS-conditional diffusion model

[0082] The forward process of the DAS-conditional diffusion model is as follows: Figure 2 As shown by the dashed lines, the mean and variance of various types of DAS noise are introduced. During the forward pass, parameterized noise conforming to the statistical characteristics of DAS noise is continuously added to the original data. The conditional distribution of noise addition transformation is performed at each time step. for:

[0083] ;

[0084] Where, x t x t-1 Let these represent the state variables at step t and step (t-1) of the diffusion process, respectively. It follows a mean of 0 and a variance of σ. 2 Gaussian distributed random noise, β t This is the noise scheduling parameter at step t, used to control the degree of noise addition at each step, and it is smoothed by a cosine-increasing form:

[0085] ;

[0086] here, It is the minimum value of the noise scheduling parameter. is the maximum value of the noise scheduling parameter, T is the total number of time steps in the diffusion process, and t is the current time step; where 1≤t≤T;

[0087] Each step of the diffusion process starts from... The mean, The process of sampling noise from a Gaussian distribution of variance and adding it to the data from the previous step, based on the above formula, effectively transforms the clean sample data into another type of noise interpolation step by step. All are derived from the output of the previous step. The result is obtained by weighted summation with noise, from to any time step The conditional distribution is:

[0088] ;

[0089] in, This is the initial sample for the diffusion process; , , representing the cumulative coefficient at different time steps during the diffusion process, used to attenuate the original sample. The influence of DAS-conditional diffusion model The noise is generated based on the statistical characteristics of DAS data, and after a sufficient time step, The process eventually converges to the prior distribution, yielding the final noisy state of the forward process. The importance of this process lies in the fact that it provides a target for the learning of the reverse denoising process, namely, it clarifies the distribution of the data to be generated;

[0090] 2) Design the reverse process of the DAS-conditional diffusion model

[0091] The reverse process of the DAS-conditional diffusion model is as follows: Figure 2 As shown by the solid line in the diagram, it is also defined on a Markov chain, which cancels the forward diffusion process and performs iterative denoising, with the goal of resolving the noise from the final noisy state. Data distribution recovered , This refers to the generated data obtained from the final modeling of the diffusion model. In the reverse process, the spectral characteristics of the noise samples obtained through short-time Fourier transform are used as condition y to guide the generation of the target data. Under the guidance of y, the reverse process obtains a more accurate target distribution. ;

[0092] A neural network is used to fit the back transition probability. :

[0093] ;

[0094] in, The variance function is fixed, and its value is... , The mean function of the network's learning predictions.

[0095] ;

[0096] Here It is the standardized residual predicted by the neural network. By training the network, it learns how to predict noise components at a specific time step. In order to effectively generate DAS noise data that conforms to physical laws, an improved U-Net network is adopted. The overall network architecture of the improved U-Net network is as follows. Figure 3 As shown, its input terminal includes noisy signal samples. The noise prior information y, which serves as a condition, is initially fused by a 3×3 convolutional layer and then enters multiple residual blocks. The residual blocks extract and map the input features step by step and connect to subsequent modules through feature propagation.

[0097] In the middle section, a temporal embedding module is included. This module converts the time step t during the diffusion process into a corresponding embedding representation and applies it to the computation of the residual block, thereby adapting to different diffusion stages. Simultaneously, an intermediate attention module is placed at the bottleneck position of the network to model global dependencies at the stage where features are most concentrated, thus improving the ability to capture noisy structures. The specific structure of the intermediate attention module is as follows... Figure 4 As shown, attention weights are generated by calculating the correlation between the query matrix Q, the key matrix K, and the value matrix V, thereby optimizing feature representation. Specifically, the calculation process of self-attention is as follows:

[0098] ;

[0099] in, To query the dimensions of the key matrix and the key matrix, The transpose of the key matrix K is used by this module to enable the network to identify and utilize long-range dependencies in DAS data, thereby improving the physical consistency of generated noise.

[0100] The U-Net network achieves multi-scale feature fusion and reconstruction through symmetrically arranged residual blocks combined with bottom-up skip connections. Finally, the decoder output is processed by the output module to generate the noise component predicted by the network. , to be used in the reverse reconstruction process of the diffusion model;

[0101] During the generation phase, the diffusion model uses an improved U-Net network to progressively eliminate noise in the data, generating data step by step. The optimization function of this network is as follows:

[0102] ;

[0103] After training, the network will master the Gaussian transition probabilities during the reverse process. ,from The generated data is obtained by sampling. :

[0104] ;

[0105] in, The preset reverse process noise scale, It follows a mean of 0 and a variance of . The random variables are Gaussian distributed. Finally, the DAS-conditional diffusion model generates noise-free data step by step through sampling steps. :

[0106] .

[0107] (3) Model training and data generation.

[0108] 1) Training the DAS - Conditional Diffusion Model

[0109] Based on the constructed model, the model is trained, and multi-type DAS noise data is generated based on the trained model. The specific process is as follows:

[0110] First, the four types of DAS noise samples extracted separately and their corresponding noise category labels are used as input data pairs to construct training samples, resulting in four noise training sets. The four noise training sets are divided into training and validation sets in a 7:3 ratio. Each sample is associated with corresponding statistical features, including mean, variance, and spectral feature matrix extracted through short-time Fourier transform.

[0111] For each training step, a time step t is randomly selected, and Gaussian noise related to the time step is added to the original noise sample according to the diffusion process formula, thereby obtaining the noisy sample at that time step. Then, the noisy samples The conditional information y and time step t are input into the U-Net network structure of the conditional diffusion model. The network outputs the corresponding noise prediction result. The loss function is calculated between the prediction result and the known real noise, and the network parameters are updated through backpropagation. Through iterative training with a large number of samples, the model can gradually learn the statistical distribution and category characteristics of DAS noise. The mean absolute error (MAE) is used as the loss function to measure the difference between the generated noise and the real noise.

[0112] ;

[0113] 2) DAS noise data generation

[0114] After the model training is completed, noise data is generated. First, random Gaussian noise is used as the initial input, and the conditional information y of the target noise is given. Then, the model iterates from the preset maximum time step T to time step 0 according to the back diffusion sampling process. In each iteration, the model receives the current noisy sample and conditional information, predicts the noise component, and calculates the sample of the previous time step according to the back diffusion formula. After multiple iterations, a synthetic DAS noise sample consistent with the target category features is finally obtained.

[0115] The effects of the present invention will be further illustrated below through specific experimental examples.

[0116] This experimental example combines the noise characteristics and data processing requirements of a real-world DAS system. Based on the method of this invention, a noise model is constructed and data is generated to verify its practical effectiveness in training set expansion and denoising tasks. The specific steps are as follows:

[0117] (1) Acquire and analyze DAS noise data;

[0118] 1) Collect DAS noise data;

[0119] In this experimental example, noise data from the DAS system was first acquired and preprocessed. Data was obtained from actual DAS records from boreholes in the Tarim Basin. Thirty representative records were selected to form the original acquisition library. Background intervals before the arrival of vibration were identified and extracted from each record as candidate segments for pure noise. If the records contained obvious earthquake events or human-induced vibrations, the time periods containing signals were removed, and pure noise samples containing only various typical mixed noises and no earthquake signal components were extracted.

[0120] 2) Extract four typical noise types from pure noise samples;

[0121] In the acquired pure noise samples, bandpass filtering was applied to separate and extract four typical noise samples. Each noise sample was then cropped into 1000 training segments with a fixed block size of 256×256 as the model input unit. The cropping was performed according to a time window (fixed duration) to finally obtain the four noise training sets.

[0122] 3) Analyze the statistical characteristics of noise;

[0123] Normalization was performed on each type of noise, and statistical characteristic analysis was conducted separately. The mean and variance of each type of noise were calculated to parameterize the noise during the forward diffusion process. Table 1 shows the mean and variance of the four types of noise (the data have been normalized and are dimensionless).

[0124] Table 1. Mean and variance of various types of noise extracted from a real DAS system

[0125] Random noise Fading noise Horizontal noise Optical noise mean <![CDATA[-3.03×10 -2 ]]> <![CDATA[2.35 × 10 -1 ]]> <![CDATA[-7.98 × 10 -2 ]]> <![CDATA[-2.70 × 10 -2 ]]> variance <![CDATA[1.78 × 10 -3 ]]> <![CDATA[7.53 × 10 -1 ]]> <![CDATA[3.02 × 10 -1 ]]> <![CDATA[3.45 × 10 -2 ]]>

[0126] On the other hand, the spectral characteristics of various noises are obtained by short-time Fourier transform as a constraint on the consistency between the generated samples and the real samples of the diffusion model.

[0127] (2) Constructing the DAS-Conditional Diffusion Model

[0128] 1) Design the forward process of the DAS - Conditional Diffusion Model

[0129] The four typical DAS noise training sets extracted above are selected, and amplitude normalization is performed to ensure that the amplitude distribution is within the interval [-1, 1]. The time step of the forward diffusion process is set to T = 1000, and the diffusion intensity coefficient is set to... Minimum value of noise scheduling parameter using cosine increment method Set to 1×10 -4 The maximum value of the noise scheduling parameter Set in the interval 2×10 -2 ;

[0130] 2) Design the inverse process of the DAS - conditional diffusion model

[0131] The model employs an improved U-Net as the core noise predictor for backpropagation. The U-Net network has a baseline of 64 channels and consists of four downsampling layers and four upsampling layers. A self-attention module is added at the bottom layer. Residual blocks are all implemented using 3×3 convolutions + GroupNorm + Swish activation. The temporal embedding dimension is set to 256. The time step t is mapped to a high-dimensional vector through positional encoding and injected into each residual module to achieve temporal-dependent modeling. A self-attention module is introduced in the middle of the network to calculate Q / K / V and normalize the weights, thereby capturing the long-range correlation of noise in time and space and enhancing the modeling ability of complex textures (such as checkerboard / optical stripes). The output module uses a GroupNorm + Swish activation + 3×3 convolution to restore the noise prediction to the same scale as the input. The backpropagation process also has a time step of T=1000. In each iteration, the model predicts the noise component and corrects it, finally obtaining the generated sample.

[0132] (3) Model training and data generation;

[0133] 1) Train the DAS - Conditional Diffusion Model;

[0134] The four types of DAS noise samples extracted separately and their corresponding noise category labels were used as input data pairs to construct training samples. A total of four noise training sets were constructed. The four noise training sets were divided into training and validation sets in a 7:3 ratio. Each sample was associated with corresponding statistical features, including mean, variance, and spectral feature matrix extracted by short-time Fourier transform.

[0135] For each training step, a time step t is randomly selected, and Gaussian noise related to the time step is added to the original noise sample according to the diffusion process formula, thereby obtaining the noisy sample at that time step. Then, the noisy samples The conditional information y and the time step t are input into the U-Net network structure of the conditional diffusion model, and the network outputs the corresponding noise prediction result.

[0136] During training, the optimizer used was Adam, and the learning rate was set to 2×10. -4 The batch size was 16, the total number of training rounds was 500, and the loss function was MAE to constrain the difference between the model's predicted noise and the actual noise. During training, the model iteratively updated the weight parameters and gradually learned the statistical distribution characteristics of different types of DAS noise at each time step. After training, the model can accurately predict the noise component in noisy samples based on the input category conditions and time step information. The DAS-conditional diffusion model was trained using four types of noise training sets to obtain four training models. After training, each model can generate new samples for the corresponding noise.

[0137] 2) Generation of DAS noise data;

[0138] After model training, data is generated for four types of noise. First, random Gaussian noise of size 256×256 is used as the initial input, along with the conditional information y of the target noise. Then, the model iterates 1000 steps in reverse order according to the backdiffusion sampling process, progressing to time step 0. In each iteration, the model receives the current noisy sample and the conditional information, and predicts the noise components. The sample from the previous time step was calculated based on the back diffusion formula. After multiple iterations, synthetic DAS noise samples consistent with the target category features were finally obtained. Ultimately, 4000 samples were generated for each noise category, forming an expanded noise training set to verify the effectiveness of the generated noise samples. Some of the generated noise samples are shown below. Figure 5 As shown.

[0139] In this experimental example, to verify the consistency between the generated data and the real data, a comparative analysis of the power spectral density and data distribution was performed on the original samples and generated samples of four types of noise. For example... Figure 6As shown in the figure, the comparison results of the four types of noise in terms of power spectral density are presented; as... Figure 7 As shown in the figure, the comparison results of the two on the data distribution map are presented. The comparison results show that the generated sample and the real sample exhibit a high degree of consistency in terms of main wave characteristics, amplitude range, distribution pattern, and power spectrum characteristics. This indicates that the noise data generated by this invention can truly reflect the statistical characteristics of DAS noise, thus verifying the authenticity and validity of the generated data.

[0140] In this experimental example, to further verify the role of the generated samples in downstream tasks, the noisy data generated by this invention was added to the original small sample training set to construct an expanded training set. A comparative experiment was then conducted using the DnCNN network as the denoising model. During training, the DnCNN network was trained using both the training set containing only the original small samples and the expanded training set containing the generated data, with all training parameters kept consistent: the learning rate was set to 1×10⁻⁶. -4 The batch size was 16, the training epochs were 200, and the loss function was the mean squared error (MSE). Under the same conditions, two denoising models were finally obtained for downstream denoising comparison and analysis.

[0141] The denoising results are shown in Figure 8: The model trained on only a small number of samples has a certain suppression effect on low-frequency noise, but high-frequency residues are significant, and signal edge details are blurred, indicating insufficient signal edge preservation ability; The model trained with the extended training set generated by this invention achieves a better balance between high-frequency noise suppression and signal edge detail preservation, and can better preserve key signal features such as continuous time travel and weak reflection events, while significantly reducing background artifacts in the residual map.

[0142] Experiment Summary:

[0143] This invention's diffusion model progressively adds and removes noise from data via a Markov chain, enabling accurate modeling of the complex distribution of a particular data type. Leveraging the advantages of diffusion models, it effectively models the noise distribution by gradually introducing noise into DAS noise data during the forward pass and restoring the data during the reverse pass. Unlike traditional diffusion models, this invention incorporates prior features of DAS noise as conditions into the model, combining the spatiotemporal correlation and multi-scale characteristics of DAS noise for modeling. This design allows the model to more accurately capture the complex distribution characteristics of DAS noise under different environments, improves the physical consistency of the generated data, facilitates the processing of various types of DAS noise, adapts to different application scenarios, and possesses strong generalization capabilities.

[0144] Through the above application verification, the noise data generated by the present invention is highly consistent with the real noise in the time-frequency domain and statistical characteristics. It can effectively expand the training set and significantly improve the performance of the DAS denoising network, especially suitable for complex noise environments and small sample scenarios.

[0145] This invention addresses the prevalent multi-type, multi-scale noise problem in DAS systems within complex geological and engineering environments by proposing a DAS-conditional diffusion model for DAS noise modeling and generation. This method fully utilizes the statistical characteristics and category priors of measured DAS noise, introducing a parameterized noise injection mechanism during the forward diffusion process and guiding Markov chain sampling through conditional noise samples during the reverse generation process. This ensures that the generated noise is highly consistent with the real noise in terms of time-domain waveform, probability distribution, and spectral characteristics.

[0146] This invention effectively solves the problems of insufficient accuracy in DAS noise modeling and poor physical consistency of generated data through conditional diffusion modeling technology without relying on large-scale labeled data. It provides an efficient, controllable, and scalable technical solution for improving DAS data quality in complex environments. This method not only enriches the means of DAS noise generation modeling in theory but also demonstrates significant application value and promising prospects in engineering practice.

Claims

1. A DAS noise modeling method based on a conditional diffusion model, characterized in that, Includes the following steps: (1) Acquire and analyze DAS noise data; (2) Constructing the DAS-conditional diffusion model; (3) Model training and data generation.

2. The DAS noise modeling method based on the conditional diffusion model according to claim 1, characterized in that: Step (1) includes: 1) Collect DAS noise data; Collect noisy data from the actual DAS system, select background noise and passive data before the arrival of the earthquake, and extract pure noise samples containing only various typical mixed noises and no seismic signal components to avoid the seismic signal components from interfering with subsequent analysis; 2) Extract four types of typical noise from pure noise samples; Based on the obtained pure noise samples, four types of typical DAS noise are separated and extracted by bandpass filtering. At the same time, the four types of noise are normalized and cropped into data blocks of size 256×256 to obtain four types of noise training sets. 3) Analyze the statistical characteristics of noise; Perform statistical characteristic analysis on the four typical noises separated, calculate the mean μ and variance σ² of each type of noise, and obtain the spectral characteristic information of each type of noise through short-time Fourier transform to provide a basis for the subsequent design of diffusion model parameters.

3. The DAS noise modeling method based on the conditional diffusion model according to claim 1, characterized in that: Step (2) includes: 1) Design the forward process of the DAS-conditional diffusion model; The mean and variance of various types of DAS noise are introduced, and parameterized noise that conforms to the statistical characteristics of DAS noise is continuously added to the original data during the forward process. The conditional distribution of noise addition transformation is performed at each time step. for: ； Where, x t x t-1 Let these represent the state variables at step t and step (t-1) of the diffusion process, respectively. It follows a mean of 0 and a variance of σ. 2 Gaussian distributed random noise, β t This is the noise scheduling parameter at step t, used to control the degree of noise addition at each step, and it is smoothed by a cosine-increasing form: ； here, It is the minimum value of the noise scheduling parameter. is the maximum value of the noise scheduling parameter, T is the total number of time steps in the diffusion process, and t is the current time step; where 1≤t≤T; Each step of the diffusion process starts from... The mean, The process of sampling noise from a Gaussian distribution of variance and adding it to the data from the previous step, based on the above formula, effectively transforms the clean sample data into another type of noise interpolation step by step. All are derived from the output of the previous step. The result is obtained by weighted summation with noise, from to any time step The conditional distribution is: ； in, This is the initial sample for the diffusion process; , , representing the cumulative coefficient at different time steps during the diffusion process, used to attenuate the original sample. The influence of DAS-conditional diffusion model The noise is generated based on the statistical characteristics of DAS data, and after a sufficient time step, The process eventually converges to the prior distribution, yielding the final noisy state of the forward process. ; 2) Design the reverse process of the DAS-conditional diffusion model Defined on a Markov chain, which undoes the forward diffusion process and performs iterative denoising, its goal is to denoise from the final noisy state. Data distribution recovered , This refers to the generated data obtained from the final modeling of the diffusion model. In the reverse process, the spectral characteristics of the noise samples obtained through short-time Fourier transform are used as condition y to guide the generation of the target data. Under the guidance of y, the reverse process obtains a more accurate target distribution. ; A neural network is used to fit the back transition probability. : ； in, The variance function is fixed, and its value is... , The mean function of the network's learning predictions. ； Here This refers to the standardized residuals predicted by the neural network. The network is trained to predict noise components at specific time steps. To effectively generate DAS noise data that conforms to physical laws, an improved U-Net network is used. The U-Net network achieves multi-scale feature fusion and reconstruction through symmetrically arranged residual blocks combined with bottom-up skip connections. Finally, the decoder output is processed by the output module to generate the noise components predicted by the network. , to be used in the reverse reconstruction process of the diffusion model; During the generation phase, the diffusion model uses an improved U-Net network to progressively eliminate noise in the data, generating data step by step. The optimization function of this network is as follows: ； After training, the network will master the Gaussian transition probabilities during the reverse process. ,from The generated data is obtained by sampling. : ； in, The preset reverse process noise scale, It follows a mean of 0 and a variance of . The random variables are Gaussian distributed. Finally, the DAS-conditional diffusion model generates noise-free data step by step through sampling steps. : 。 4. The DAS noise modeling method based on the conditional diffusion model according to claim 3, characterized in that: The improved U-Net network in step 2) includes noisy signal samples at its input. The noise prior information y, which serves as a condition, is initially fused by a 3×3 convolutional layer and then enters multiple residual blocks. The residual blocks extract and map the input features step by step and connect to subsequent modules through feature propagation. In the middle section, a temporal embedding module is set up. This temporal embedding module converts the time step t in the diffusion process into the corresponding embedding representation and applies it to the operation of the residual block to achieve adaptation to different diffusion stages. At the same time, an intermediate attention module is set up at the bottleneck position of the network to model global dependencies in the stage where features are most concentrated, thereby improving the ability to capture noise structures. Through this module, the network can identify and utilize long-distance dependencies in DAS data to improve the physical consistency of generated noise.

5. The DAS noise modeling method based on the conditional diffusion model according to claim 4, characterized in that: The intermediate attention module generates attention weights by calculating the correlation between the query matrix Q, the key matrix K, and the value matrix V, thereby optimizing feature representation; the calculation process of self-attention is as follows: ； in, To query the dimensions of the key matrix and the key matrix, Let K be the transpose of the key matrix K.

6. The DAS noise modeling method based on the conditional diffusion model according to claim 1, characterized in that: Step (3) includes: 1) Train the DAS-conditional diffusion model; Based on the constructed model, the model is trained, and multi-type DAS noise data is generated based on the trained model. 2) Generation of DAS noise data; A synthetic DAS noise sample consistent with the target category features is obtained.

7. The DAS noise modeling method based on the conditional diffusion model according to claim 6, characterized in that: The specific training process for step 1) is as follows: First, the four types of DAS noise samples extracted separately and their corresponding noise category labels are used as input data pairs to construct training samples, resulting in four noise training sets. The four noise training sets are divided into training and validation sets in a 7:3 ratio. Each sample is associated with corresponding statistical features, including mean, variance, and spectral feature matrix extracted through short-time Fourier transform. For each training step, a time step t is randomly selected, and Gaussian noise related to the time step is added to the original noise sample according to the diffusion process formula, thereby obtaining the noisy sample at that time step. Then, the noisy samples The conditional information y and time step t are input into the U-Net network structure of the conditional diffusion model. The network outputs the corresponding noise prediction result. The loss function is calculated between the prediction result and the known real noise, and the network parameters are updated through backpropagation. Through iterative training with a large number of samples, the model can gradually learn the statistical distribution and category characteristics of DAS noise. The mean absolute error (MAE) is used as the loss function to measure the difference between the generated noise and the real noise. 。 8. The DAS noise modeling method based on the conditional diffusion model according to claim 6, characterized in that: Step 2) includes generating noise data after the model training is completed. First, random Gaussian noise is used as the initial input, and the conditional information y of the target noise is given. Then, the model iterates from the preset maximum time step T to time step 0 according to the back diffusion sampling process. In each iteration, the model receives the current noisy sample and conditional information, predicts the noise component, and calculates the sample of the previous time step according to the back diffusion formula. After multiple iterations, a synthetic DAS noise sample consistent with the target category features is finally obtained.

Images