A fiber optic seismic record reconstruction method based on CVAE and detector constraints

Abstract

This invention discloses a fiber optic seismic record reconstruction method based on CVAE and detector constraints, belonging to the field of geophysical exploration and signal processing technology. It utilizes discretely distributed high-precision seismometers as anchor points to segment distributed acoustic wave sensing fiber optic acquisition data. A pre-trained reconstruction model, centered on a conditional variational autoencoder (CVAE), is constructed to obtain general mapping characteristics for converting fiber optic acquisition data to high-fidelity seismic detector acquisition data. The fiber optic acquisition data is preprocessed to align its modes with those of the seismic detector acquisition data. Each data segment and the seismic detector acquisition data at its two end anchor points are used as input constraints. Leveraging the continuous evolution characteristics of the CVAE latent space and combining real-time double-ended strong boundary constraints, a high-fidelity seismic record with high sampling density is output. This effectively solves the technical challenge of integrating high-density observation with high-quality processing systems, thereby improving exploration accuracy.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration and signal processing technology, specifically a fiber optic seismic record reconstruction method based on CVAE and detector constraints. Background Technology

[0002] In seismic exploration and geological monitoring, spatial sampling rate and signal fidelity are core indicators for evaluating the performance of observation systems. Traditional seismic detectors measure particle vibration velocities, have extremely high signal-to-noise ratios and clear physical meanings, and the acquired data is relatively accurate; however, due to limitations in deployment costs and field conditions, their channel spacing is typically on the order of ten to tens of meters, resulting in sparse spatial sampling (i.e., scattered sampling points).

[0003] Distributed acoustic sensing (DAS) technology utilizes optical fiber as the sensing medium, enabling ultra-high-density spatial sampling at the meter or even centimeter level, greatly enriching wavefield information. However, the raw DAS signal is generally low in signal-to-noise ratio due to the influence of laser phase noise and uneven fiber coupling, resulting in poor performance for direct data processing, and the related algorithm system is currently immature. More importantly, DAS signals differ significantly from traditional detectors in waveform morphology and physical meaning, making it impossible to directly apply mature processing algorithms (such as filtering, migration, and positioning) accumulated in the seismic exploration field to DAS data.

[0004] Therefore, the research direction of this invention is to accurately and in real time calibrate and reconstruct a seismic record with high-fidelity characteristics of traditional geophones and high sampling density, thereby being compatible with existing mature processing systems and effectively improving exploration accuracy. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a fiber optic seismic record reconstruction method based on CVAE and detector constraints. By constructing a CVAE mapping model with double-ended constraints, the DAS signal can be accurately and reconstructed in real time into a seismic record that possesses the high fidelity characteristics of traditional detectors and maintains high sampling density, effectively improving exploration accuracy.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a fiber optic seismic record reconstruction method based on CVAE and detector constraints, comprising the following steps: Step 1: Deploy the observation system: Deploy distributed acoustic sensing fiber (DAS) in the target monitoring area, and deploy multiple seismic detectors along the distributed acoustic sensing fiber at fixed intervals to form an observation system.

[0007] Step 2, Synchronous Acquisition and Data Segmentation: Start the observation system to sample seismic data throughout the day, and divide the data acquired by the distributed acoustic wave sensing fiber optic into data segments between adjacent seismic detectors.

[0008] Step 3: Physical alignment based on multi-order derivatives of detector type: According to the type of seismic detector used in the field, each data segment in Step 2 is preprocessed by time derivative to achieve physical alignment with the seismic detector data.

[0009] Step 4: Construct a CVAE mapping model with double-ended constraints: First, construct a mapping model with a conditional variational autoencoder (CVAE) as its core. Use each data segment processed in Step 3, as well as the seismic detector data collected at the anchor points at both ends of each data segment, as condition variables. A common input model is used; the model obtains a mapping function for converting data acquired from optical fiber to data acquired from high-fidelity seismic detectors through nonlinear mapping of the latent space.

[0010] Step 5: Model pre-training: The model constructed in Step 4 is pre-trained using a historical earthquake sample database.

[0011] Step 6: Full-space high-density wavefield reconstruction: After training is completed, repeat steps 2 and 3, inputting the seismic detector data collected by each data segment and the anchor points at both ends of each data segment into the model trained in step 5, thereby outputting a seismic record with high fidelity and maintaining high sampling density.

[0012] Furthermore, in step one, the demodulation parameters of the distributed acoustic sensing fiber are set to include the gauge length and channel spacing d; the fixed spacing is D, and D is an integer multiple of d. This ensures that each data segment after partitioning contains complete demodulation point data from multiple fibers.

[0013] Furthermore, after the deployment in step one is completed, a unified clock source is used to perform time synchronization calibration between the distributed acoustic sensing fiber optic cable and each seismic detector. This synchronization calibration ensures the synchronization of data acquired by the distributed acoustic sensing fiber optic cable and the seismic detectors, facilitating subsequent data analysis.

[0014] Furthermore, after the division is completed in step two, the data collected from all fiber demodulation points within each data segment is obtained. And the seismic detector data acquired at both ends of this data segment. and The above data will be used to construct a CVAE mapping model with double-ended constraints.

[0015] Furthermore, in step three, if the seismic detector is a velocity-type detector, the first-order time derivative of each data segment in step two is performed to obtain the strain rate signal, aligning it with the vibration velocity mode; if the seismic detector is an acceleration-type detector, the second-order time derivative of each data segment in step two is performed to obtain the time derivative signal of the strain rate, aligning it with the vibration acceleration mode. This preprocessing initially eliminates the phase and morphological differences between the DAS-acquired data and the seismic detector-acquired data, establishing a preliminary alignment of physical features.

[0016] Furthermore, the specific formula for the mapping function in step four is as follows: in, Indicates the reconstructed position Earthquake data at the location; This represents the fiber optic acquisition data after preprocessing in step three; Represents a condition vector containing information about two-sided constraints; and These represent the encoder function and the decoder function, respectively.

[0017] Furthermore, in step four, the model uses a nonlinear mapping of the latent space, specifically formulated as follows: Among them, the encoder function Compare the input data segment with the condition vector Compressed into latent variables Latent variables Capture the nonlinear feature mapping between fiber optic acquisition data and seismic detector acquisition data within the probability space; decoder Under the strong constraints of the data acquired by the seismic detectors at both ends, the latent variables will be... Reconstructed as location Earthquake data at the location This means restoring the earthquake record to a high fidelity.

[0018] Compared with existing technologies, this invention utilizes spatially discrete high-precision seismometers as anchor points to perform real-time physical mapping and quality enhancement on data acquired by high-density distributed acoustic wave sensing fiber optics. Specifically, a high-precision seismograph is placed at fixed intervals along the distributed acoustic wave sensing fiber optic path as a physical calibration point. At the data processing end, a pre-trained reconstruction model based on a conditional variational autoencoder (CVAE) is constructed. This model is first pre-trained on a historical earthquake sample database containing diverse geological features to obtain general mapping characteristics for converting data acquired from fiber optics to high-fidelity seismograph data. In practical applications, the model performs time derivative preprocessing on the DAS-acquired data according to the type of on-site seismograph (velocity or acceleration) to align its modes with those of the seismograph data. Subsequently, the seismic detector data collected at each fiber demodulation point of the location to be reconstructed and the anchor points at both ends of its data segment are used together as input constraints. Utilizing the continuous evolution characteristics of the CVAE latent space and combined with real-time double-ended strong boundary constraints, the model can automatically correct noise and distortion in the fiber signal and complete cross-modal physical mapping. Through the above process, this invention can achieve comprehensive calibration of signal quality towards high-precision seismic detectors while preserving the ultra-high spatial resolution of distributed acoustic wave sensing fibers, effectively solving the technical challenge of integrating high-density observation and high-quality processing systems, thereby improving exploration accuracy. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the layout of the observation system in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the construction of condition vectors based on double-ended anchor point constraints in an embodiment of the present invention.

[0021] Figure 3 This is a network architecture diagram for reconstructing earthquake records in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below.

[0023] This invention addresses the pain points of low signal-to-noise ratio, mismatch between physical modes and traditional geophone data, and inability to be directly adapted to mature seismic processing systems in seismic exploration using discrete high-precision geophones as anchor constraints and conditional variational autoencoders (CVAEs). The invention reconstructs DAS data into high-fidelity seismic records by using discrete high-precision geophones as anchor constraints and combining them with conditional variational autoencoders (CVAEs). The following two specific embodiments are described in detail.

[0024] Example 1:

[0025] This example illustrates a seismic exploration scenario for shallow shale gas on land. The target monitoring area is a shale gas exploration block with an exploration depth of 0-3000m. Surface wave and environmental noise interference are significant, necessitating the acquisition of seismic data with both ultra-high spatial sampling density and a high signal-to-noise ratio. Figure 1 As shown, the specific implementation steps are as follows: Step 1: Deployment of the Observation System: A single-mode G.652D communication fiber optic cable is deployed as a distributed acoustic wave sensing fiber optic cable in the target monitoring area, with a total survey line length of 2000m. An iDAS-2000 distributed acoustic wave sensing demodulator is used, with demodulation parameters set as follows: gauge length 10m, sampling frequency 1000Hz, channel spacing d=5m, meaning one fiber demodulation point is formed every 5 meters, for a total of 400 fiber demodulation points along the entire survey line. Multiple seismic detectors are deployed along the distributed acoustic wave sensing fiber optic cable at a fixed spacing D of 20 meters. In this embodiment, a 20DX velocity-type seismic detector is used. Each detector is tightly coupled to the fiber optic cable and buried at a depth of 0.3m to avoid surface wind and vibration interference, ultimately forming the observation system. This ensures that each data segment contains complete data from the five fiber optic demodulation points after partitioning. After deployment, a high-precision BeiDou-2 clock source (time accuracy ±1μs) is used to synchronize and calibrate the acquisition systems of the DAS demodulator and all seismic detectors, ensuring that the time synchronization error of the acquired data is less than 10μs, thus completely eliminating the impact of time phase deviation on subsequent data processing.

[0026] Step 2: Synchronous Acquisition and Data Segmentation: Start the observation system to sample seismic data throughout the entire time period, and divide the distributed acoustic wave sensing fiber optic data into segments based on the distance between adjacent seismic detectors; after segmentation, acquire the data from all fiber optic demodulation points within each data segment. And the seismic detector data acquired at both ends of this data segment. and The above data will be used to construct a CVAE mapping model with double-ended constraints.

[0027] Step 3: Physical Alignment Based on Detector Type and Multi-Derivative: Depending on the type of seismic detector used in the field, each data segment from Step 2 undergoes time derivative preprocessing to achieve physical alignment with the seismic detector data. In this embodiment, a velocity-type seismic detector is used. Therefore, the raw fiber strain data in each data segment of Step 2 undergoes first-order time derivative preprocessing. The derivative calculation uses the central difference method with a difference step size of one sampling interval (1ms) to obtain the strain rate signal, achieving physical mode alignment with the vibration velocity signal acquired by the detector. After preprocessing, the differentiated fiber data and the data from the anchor point detectors at both ends undergo unified bandpass filtering. The filtering frequency band is set to 3~120Hz to filter out low-frequency drift and high-frequency random noise, further eliminating phase and morphological differences between the two types of data, completing the initial physical feature alignment.

[0028] Step 4: Construct a CVAE mapping model with dual-ended constraints: First, construct a mapping model with a conditional variational autoencoder (CVAE) as its core. The model consists of three parts: an encoder, a latent space layer, and a decoder. Dual-ended anchor data is introduced as conditional constraint variables. The complete network architecture is as follows: Figure 3 As shown.

[0029] The data from each data segment processed in step three, as well as the seismic detectors at the anchor points at both ends of each data segment, will be collected. and As a condition variable The common input model; the model obtains a mapping function for converting data acquired from fiber optic acquisition to data acquired by a high-fidelity seismograph through a nonlinear mapping of the latent space, specifically: The specific formula for the mapping function is: in, Indicates the reconstructed position Earthquake data at the location; This represents the fiber optic acquisition data after preprocessing (differentiation and alignment) in step three. Represents a condition vector containing information about two-sided constraints; and These represent the encoder function and the decoder function, respectively.

[0030] like Figure 2 As shown, the above condition vector The specific formula is: in, and Data was collected from the seismic detectors at both ends of each data segment. Location to be reconstructed The normalized spatial weight scalar relative to the two anchor points is calculated as follows: ; and These are the positions of the two anchor points; A time-domain feature extraction operator is used to extract waveform features from the seismic detector data acquired at the double-ended anchor points using a one-dimensional convolutional neural network. As a spatial embedding operator, a fully connected layer maps the position scalar into high-dimensional features to represent the relative spatial relationship between the reconstruction and the two end anchor points; This is a feature concatenation operation.

[0031] The model uses a nonlinear mapping of the latent space, specifically formulated as follows: The encoder employs a 4-layer one-dimensional convolution and pooling structure, and the encoder function... Compare the input data segment with the condition vector Compressed into latent variables Latent variables Capture the nonlinear feature mapping between fiber optic acquisition data and seismic detector acquisition data within the probability space; decoder Under the strong constraints of the data acquired by the seismic detectors at both ends, the latent variables will be... Reconstructed as location Earthquake data at the location This means restoring the earthquake record to a high fidelity.

[0032] Step 5: Model Pre-training: The model constructed in Step 4 is pre-trained using a historical earthquake sample database for the region. This database contains 120 sets of DAS and velocity detector data collected synchronously along the same seismic line, covering different geological structures, source types, and signal-to-noise ratio scenarios. 96 sets are used as the training set, and 24 sets as the validation set. The pre-training hardware environment uses an NVIDIA RTX 4090 GPU, and the software framework uses PyTorch 2.0 and Python 3.10. The training parameters are set as follows: the optimizer is AdamW, the initial learning rate is 1e-4, the weight decay is 1e-5, the batch size is 8, and the number of training epochs is 200. When the validation set loss no longer decreases for 10 consecutive epochs, an early stopping mechanism is triggered to save the optimal model weights.

[0033] To further improve model accuracy, after pre-training, during the field application phase, the model was dynamically fine-tuned 50 times using real-time data from 100 excitation sources simultaneously collected at 11 anchor points in this embodiment. This further reduced the model's reconstruction error at the anchor points, allowing the model parameters to automatically adapt to the geological environment, fiber optic coupling status, and acquisition system errors of the current block. The learning rate was set to 1e-5 during the fine-tuning phase, while the remaining training parameters remained unchanged.

[0034] Step Six: Full-Space High-Density Wavefield Reconstruction: After training, repeat steps two and three, inputting the seismic detector data collected at each data segment and the anchor points at both ends of each data segment into the model trained in step five. Based on the nonlinear mapping process of the latent space, the model performs sliding inference on the data of each fiber optic detuning point in each data segment, thereby outputting seismic records with high fidelity and maintaining high sampling density. Then, the seismic records obtained from each data segment are used as a set, and finally a high-density seismic gather with ultra-high spatial resolution and adapted to mature algorithm systems is synthesized.

[0035] The reconstructed seismic record obtained in this embodiment, compared with the real data collected by the verification detector (not involved in model constraints and training) deployed at an additional location 1000m along the survey line, showed a waveform correlation coefficient of 0.942 and a signal-to-noise ratio improvement of 28.6dB over the original DAS data. This effectively eliminated the phase noise and amplitude distortion of the original DAS data, achieving the integration of high-density observation and high-quality processing systems.

[0036] Example 2:

[0037] This embodiment shares the same core technical solution as Embodiment 1, with the only difference being the detector type and corresponding preprocessing method, as detailed below: 1. In step one, the seismic detector used is the JZ-20 type acceleration seismic detector, and the number, spacing and coupling method are the same as in Example 1.

[0038] 2. In step three, for the acceleration detector, the original fiber strain data in each data segment is preprocessed by second-order time derivative to obtain the time derivative signal of the strain rate, so that it is physically aligned with the vibration acceleration signal collected by the detector; the second-order derivative is performed by the second-order central difference method with a difference step size of 1ms.

[0039] 3. In step five, the pre-training dataset uses a historical earthquake sample library that includes synchronously acquired data from accelerometers. The rest of the model structure, training parameters, and data processing flow are completely consistent with Example 1.

[0040] The reconstructed acceleration seismic record obtained in this embodiment, compared with the real data collected by the calibrated geophone at the same location, has a waveform correlation coefficient of 0.927 and a signal-to-noise ratio improvement of 26.3 dB compared with the original DAS data. This achieves accurate reconstruction of DAS data into high-fidelity acceleration seismic records and can be directly adapted to the seismic processing algorithm system of acceleration geophone data.

Claims

1. A fiber optic seismic record reconstruction method based on CVAE and detector constraints, characterized in that, Includes the following steps: Step 1: Deploy distributed acoustic wave sensing optical fibers in the target monitoring area, and deploy multiple seismic detectors along the distributed acoustic wave sensing optical fibers at fixed intervals to form an observation system. Step 2: Start the observation system to sample seismic data throughout the day, and divide the data acquired by the distributed acoustic wave sensing fiber optic into a data segment between adjacent seismic detectors; Step 3: Based on the type of seismic detector used on site, perform time derivative preprocessing on each data segment from Step 2 to achieve physical alignment with the seismic detector data. Step 4: First, construct a mapping model with a conditional variational autoencoder as the core. Input each data segment processed in Step 3, as well as the seismic detector data collected at the anchor points at both ends of each data segment, as conditional variables into the model. The model obtains the mapping function that transforms the data collected from optical fiber to the data collected by the seismic detector through nonlinear mapping of the latent space. Step 5: Pre-train the model constructed in Step 4 using a historical earthquake sample database; Step 6: After completing the training, repeat steps 2 and 3, inputting the data collected by the seismic detectors at each data segment and the anchor points at both ends of each data segment into the model trained in step 5, thereby outputting the reconstructed seismic record.

2. The fiber optic seismic record reconstruction method based on CVAE and detector constraints according to claim 1, characterized in that, In step one, the demodulation parameters of the distributed acoustic wave sensing fiber are set to include gauge length and channel spacing d; the fixed spacing is D, and D is an integer multiple of d.

3. The fiber optic seismic record reconstruction method based on CVAE and detector constraints according to claim 1, characterized in that, After the deployment in step one is completed, a unified clock source is used to perform time synchronization calibration between the distributed acoustic wave sensing fiber and each seismic detector.

4. The fiber optic seismic record reconstruction method based on CVAE and detector constraints according to claim 1, characterized in that, After the division is completed in step two, the data acquisition data of all fiber optic demodulation points in each data segment, as well as the data acquisition data of the seismic detectors at both ends of the data segment, are obtained.

5. The fiber optic seismic record reconstruction method based on CVAE and detector constraints according to claim 1, characterized in that, In step three, if the seismic detector is a velocity detector, then the first-order time derivative of each data segment in step two is performed to obtain the strain rate signal, which is then aligned with the vibration velocity mode; if the seismic detector is an acceleration detector, then the second-order time derivative of each data segment in step two is performed to obtain the time derivative signal of the strain rate, which is then aligned with the vibration acceleration mode.

6. The fiber optic seismic record reconstruction method based on CVAE and detector constraints according to claim 1, characterized in that, The specific formula for the mapping function in step four is as follows: in, Indicates the reconstructed position Earthquake data at the location; This represents the fiber optic acquisition data after preprocessing in step three; Represents a condition vector containing information about two-sided constraints; and These represent the encoder function and the decoder function, respectively.

7. The fiber optic seismic record reconstruction method based on CVAE and detector constraints according to claim 6, characterized in that, In step four, the model uses a nonlinear mapping of the latent space, specifically formulated as follows: Among them, the encoder function Compare the input data segment with the condition vector Compressed into latent variables Latent variables Capture the nonlinear feature mapping between fiber optic acquisition data and seismic detector acquisition data within the probability space; decoder Under the strong constraints of the data acquired by the seismic detectors at both ends, the latent variables will be... Reconstructed as location Earthquake data at the location .

Images