Lightweight DAS seismic phase pickup method and system based on sparse connection

Abstract

The invention discloses a lightweight DAS seismic phase pickup method and system based on sparse connection, and relates to the technical field of seismic exploration, and the method comprises the steps: carrying out the segmentation and local time window division of DAS data of a target region, and dynamically adjusting a morphological structure element based on each local time window, thereby generating an enhanced signal feature sequence; the sequence is input into a parallel fusion network formed by convolution and an attention mechanism and a sparse connection U-shaped network at the same time, multi-channel feature maps output by the two types of networks are fused, and P-wave and S-wave arrival distribution curves reflecting time and space distribution are formed through corresponding matching and weighted stacking of local and global features; network redundancy is effectively reduced through a sparse connection mechanism, calculation efficiency is improved, meanwhile, in combination with local convolution response and global attention information, accurate capture of energy concentration positions and propagation trends of P waves and S waves is achieved, and meanwhile it is guaranteed that the model completely characterizes a key vibration mode under the condition of low calculation amount.

Description

Technical Field

[0001] This invention relates to the field of seismic detection technology, and in particular to a lightweight DAS phase picking method and system based on sparse connectivity. Background Technology

[0002] In modern seismic exploration and geological monitoring, distributed optical fiber acoustic sensing (DAS) technology has attracted widespread attention due to its ability to acquire large-scale, high-density, and continuous vibration signals along optical fiber lines. DAS systems can transform ordinary optical fibers into high-density sensor arrays, acquiring acoustic or vibration information along the line by demodulating phase or frequency changes, providing abundant data for seismic wave arrival detection, geological structure analysis, and monitoring of surface and subsurface activity. However, DAS signals are often accompanied by high noise backgrounds, non-ideal fiber responses, and signal attenuation issues. Furthermore, the data volume is extremely large. Traditional phase picking methods based on fully connected deep neural networks or complex convolutional networks are computationally expensive and require large storage when processing large-scale DAS data. They are also susceptible to redundancy features and local noise interference, leading to a decrease in the accuracy of P-wave and S-wave arrival identification.

[0003] In this context, DAS data acquired along optical fibers typically exhibits characteristics of localized energy concentration, significant phase fluctuations, and uneven spatial distribution, with substantial differences in signal amplitude across different channels and time periods. External interference, variations in geological conditions, and differences in fiber optic cable laying status also contribute to signal non-uniformity, making it difficult for traditional signal analysis methods to accurately identify the arrival times and spatial distributions of P-waves and S-waves. Furthermore, existing methods suffer from high computational and storage costs when processing large-scale DAS data, are sensitive to noise and local anomalies, and struggle to achieve real-time, efficient data processing while maintaining accuracy. Therefore, in DAS fiber optic cable monitoring, it is necessary to fully consider the local and global characteristics of the signal, complex propagation effects, and the impact of fiber optic cable laying methods. Appropriate preprocessing and feature analysis of DAS data are crucial to address waveform complexity and interference issues, thus providing a technical foundation for high-precision, high-reliability phase identification and waveform analysis.

[0004] For example, announcement number CN113848587B discloses a method for seismic phase arrival time picking based on a spatiotemporal attention mechanism, belonging to the fields of seismic signal processing and artificial intelligence technology. The technical problem to be solved is to provide an improvement to the seismic phase arrival time picking method based on a spatiotemporal attention mechanism. The technical solution adopted to solve the above technical problem is as follows: acquire seismic signal data and annotate the P-wave and S-wave arrival times of the data; preprocess the seismic signal data, including data augmentation, dataset segmentation, and spectral mapping; construct a sequence processing network based on the U-Net model; integrate the spatiotemporal attention mechanism into the picking model; use a deep coding feature fusion mechanism to supplement missing feature information; adjust the model parameters according to the loss value and various evaluation indicators to complete the final model construction; input the seismic signal data to be identified into the microseismic phase picking model to obtain the seismic phase arrival time picking and annotation results.

[0005] For example, CN112799128A discloses a method for seismic signal detection and phase extraction, used in an edge device-based seismic detection system. The edge device is implemented using a Jetson Nano chip, and a lightweight deep learning model, LCANet, is deployed on it. Seismic waveform data acquired at the device end is input into the edge device, which outputs seismic signal time series, P-wave phases, and S-wave phases in real time. The LCANet model extracts feature vector sequences describing the inherent physical meaning of the seismic data from the input seismic waveform data using an encoder based on inverse bottleneck residual blocks. Then, a context-aware attention module obtains feature vector sequences of the attention time series context information for three tasks. Finally, a multi-scale heterogeneous decoder maps the feature vectors to the feature space of the corresponding task.

[0006] Existing seismic phase picking techniques are generally based on deep learning models. They achieve accurate identification of P-wave and S-wave arrival times by preprocessing seismic signals, extracting features, and fusing them using spatiotemporal attention mechanisms. For example, relying on U-Net or lightweight convolutional networks, encoders extract multi-scale features of the signal and enhance these features using attention mechanisms or context-aware modules to improve the model's ability to capture seismic phases. Furthermore, these techniques emphasize real-time processing of seismic waveforms on edge computing or lightweight devices, using inverse residual blocks and multi-scale decoders to achieve feature mapping and phase identification. All these methods rely on the temporal and multi-scale spatial features of the signal, fusing global or local features through deep network structures to complete seismic phase picking. However, DAS-based fiber optic seismic data presents unique challenges: the signal distribution along the fiber optic cable is greatly affected by the cable's laying method, including complete burial, shallow burial, overhead, crossing bridges or tunnels, and close to roads or building exteriors; under different media and boundary conditions, seismic wave propagation exhibits complex forms, with body waves attenuating, surface waves potentially amplifying, and guided or coupled vibration modes possibly appearing. Furthermore, the local energy distribution and vibration modes of the signal exhibit significant non-uniformity and spatiotemporal heterogeneity. This characteristic makes existing technologies reliant on fixed network structures and global features susceptible to strong local interference and non-uniform signals, potentially leading to decreased phase picking accuracy, ineffective capture of local anomalies, and limited model computational efficiency. Moreover, existing model learning methods that rely on fixed network structures and global features result in excessive computational burden. Summary of the Invention

[0007] This invention provides a lightweight DAS phase picking method and system based on sparse connectivity. The technical solution provided by this application is as follows:

[0008] According to a first aspect of this application, a lightweight DAS phase picking method based on sparse connectivity is provided. The method includes: performing segmentation analysis on DAS data of a target region to obtain the length of each segmentation window of the DAS data of the target region; segmenting the DAS data of the target region to form a DAS signal sequence of the target region; dividing the DAS signal sequence of the target region to obtain each local time window; and dynamically adjusting the morphological structural elements based on each local time window.

[0009] The enhanced signal feature sequence is generated based on the DAS signal of the target region after the action of morphological structural elements, and is simultaneously input into a parallel fusion network composed of convolution and attention mechanisms, as well as a sparsely connected U-shaped network.

[0010] The vibration patterns of the target region are accurately characterized and identified by a parallel fusion network composed of convolution and attention mechanisms.

[0011] In the convolutional branch of the sparsely connected U-shaped network, multi-channel feature maps corresponding to each convolution and downsampling layer are constructed and monitored. At the same time, channel activation intensity analysis is performed on the multi-channel feature maps corresponding to each convolution and downsampling layer, and the execution of convolution and downsampling operations is judged.

[0012] The multi-channel feature maps output by the parallel fusion network composed of convolution and attention mechanisms and the sparse connection U-shaped network are fused to determine the arrival candidate regions of P-waves and S-waves, forming the arrival distribution curves of P-waves and S-waves.

[0013] According to another aspect of this application, a lightweight DAS phase acquisition system based on sparse connectivity is provided, including: a DAS signal segmentation and local time window generation module, used to segment and analyze DAS data of a target region, obtain the length of each segmentation window of the DAS data of the target region, segment the DAS data of the target region to form a DAS signal sequence of the target region, divide the DAS signal sequence of the target region to obtain each local time window, and dynamically adjust the morphological structural elements based on each local time window.

[0014] An enhanced signal feature sequence generation and dual-network input module is used to generate an enhanced signal feature sequence based on the DAS signal of the target region after the action of morphological structuring elements, and simultaneously input it into a parallel fusion network composed of convolution and attention mechanisms and a sparsely connected U-shaped network.

[0015] The parallel fusion network local and global vibration mode characterization module is used to accurately characterize and identify the vibration modes of the target region based on the parallel fusion network composed of convolution and attention mechanisms.

[0016] The sparsely connected U-shaped network feature map construction and channel activation monitoring module is used to construct and monitor the multi-channel feature maps corresponding to each convolution and downsampling layer in the convolutional branch of the sparsely connected U-shaped network. At the same time, it performs channel activation intensity analysis on the multi-channel feature maps corresponding to each convolution and downsampling layer and judges the execution of convolution and downsampling operations.

[0017] The multi-channel feature fusion and determination module is used to fuse the multi-channel feature maps output by the parallel fusion network composed of convolution and attention mechanisms and the sparse connection U-shaped network to determine the arrival candidate regions of P-waves and S-waves, and form the arrival distribution curves of P-waves and S-waves.

[0018] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0019] (1) The target region DAS data is segmented and local time windows are divided, and the morphological structural elements are dynamically adjusted based on each local time window to generate an enhanced signal feature sequence. This sequence is simultaneously input into a parallel fusion network composed of convolution and attention mechanisms and a sparse connection U-shaped network. The multi-channel feature maps output by the two types of networks are fused. Through the corresponding matching and weighted superposition of local and global features, P-wave and S-wave arrival distribution curves reflecting time and space distribution are formed. The sparse connection mechanism effectively reduces network redundancy and improves computational efficiency. At the same time, by combining local convolution response and global attention information, the accurate capture of the energy concentration location and propagation trend of P-wave and S-wave is achieved, while ensuring the complete representation of key vibration modes under low computational conditions.

[0020] (2) The length of the segmentation window is dynamically determined by obtaining the first segmentation window scale adjustment factor, thereby constructing a continuous DAS signal sequence for the target region. Subsequently, the signal sequence is further divided into fine-grained local time windows, the temporal vibration response of each local window is extracted, and the second local time window scale adjustment factor is calculated. Based on this, the morphological structural elements are dynamically adjusted to achieve adaptive signal enhancement and refined characterization of local features. This method can significantly improve the local feature expression capability of the signal through multi-scale segmentation and dynamic morphological adjustment, while reducing redundant information. This enables the lightweight DAS phase picking network based on sparse connections to capture the arrival time characteristics and local energy changes of P-waves and S-waves more accurately while maintaining low computational load, providing useful technical support for efficient and accurate phase detection.

[0021] (3) First, the DAS-enhanced signal of the target region after morphological structural element processing is formed into a continuous enhanced signal feature sequence in time order and input into the convolution branch. By setting convolution kernel groups of different lengths, the signal is subjected to sliding convolution operation on multiple time scales to obtain the local feature response value of each convolution kernel at each time point. At the same time, the response value is dynamically corrected according to the local energy and waveform change data to enhance the representation ability of important local features. Then, the same enhanced signal feature sequence is input into the convolution branch of the sparse connection U-shaped network. Layer-by-layer convolution and downsampling operations are performed in the encoding path to generate multi-scale spatiotemporal feature responses and construct multi-channel feature maps of each downsampling layer. The amplitude average response of each channel feature vector is calculated along the time dimension and compared with the threshold stored in the database. Low response channels are removed in skip connections to achieve the selection of key features and the suppression of redundant information. By capturing fine-grained vibration features on the time scale of the signal through local convolutional kernel responses, and by using a sparsely connected U-shaped network to efficiently encode multi-channel global features and filter low-response channels, it is possible to accurately characterize the arrival times and vibration modes of P-waves and S-waves in the DAS signal of the target region while ensuring the model's lightweight nature. This provides efficient and accurate feature input and computational optimization for lightweight DAS phase picking.

[0022] (4) First, the multi-channel feature maps of each downsampling layer of the sparsely connected U-shaped network are statistically analyzed in terms of time and channel dimensions to obtain the comprehensive complexity index value, thereby determining whether to continue convolution and downsampling operations, so as to suppress redundant features and retain key features. Then, the multi-channel feature maps output by the parallel fusion network of convolution and attention mechanism and the sparsely connected U-shaped network are arranged in terms of time and channel dimensions respectively, and jointly form candidate P-wave and S-wave arrival regions containing time interval and channel information. Finally, the candidate intervals and their amplitude information are combined with the spatial distribution of fiber channel to construct the channel amplitude matrix in time order, and generate the three-dimensional distribution curve of P-wave and S-wave arrival, which ensures the fine capture of local vibration features and retains the global vibration mode information in the channel dimension. This achieves high-precision, low-computation P-wave and S-wave arrival detection and distribution trend characterization, providing an effective feature extraction and judgment mechanism for lightweight phase picking.

[0023] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0024] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this application. Wherein:

[0025] Figure 1This is a flowchart of a lightweight DAS phase picking method based on sparse connectivity provided in an embodiment of the present invention.

[0026] Figure 2 This is a flowchart of a lightweight DAS phase picking method based on sparse connectivity provided in an embodiment of the present invention.

[0027] Figure 3 This is an overall flowchart of the end-to-end phase picking model for DAS data provided in this embodiment of the invention;

[0028] Figure 4 This is a schematic diagram of the parallel convolution-attention feature fusion module provided in an embodiment of the present invention;

[0029] Figure 5 This is a diagram of the sparse attention module architecture based on U-Net and skip connections provided in an embodiment of the present invention;

[0030] Figure 6 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0031] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0032] This invention provides a lightweight DAS phase acquisition method based on sparse connectivity. This method can be implemented by an SSS device, which can be a terminal or a server. Figure 1 The flowchart shown is for a lightweight DAS phase picking method based on sparse connectivity. The processing flow of this method may include the following steps:

[0033] The target region DAS data is segmented and analyzed to obtain the length of each segmentation window. The target region DAS data is segmented to form a target region DAS signal sequence. The target region DAS signal sequence is divided to obtain each local time window. The morphological structural elements are dynamically adjusted based on each local time window.

[0034] An enhanced signal feature sequence is generated based on the DAS signal of the target region after the action of morphological structural elements, and simultaneously input into a parallel fusion network composed of convolution and attention mechanisms and a sparse connection U-shaped network.

[0035] The vibration patterns of the target region are accurately characterized and identified by a parallel fusion network composed of convolution and attention mechanisms.

[0036] In the convolutional branch of the sparsely connected U-shaped network, multi-channel feature maps corresponding to each convolution and downsampling layer are constructed and monitored. At the same time, channel activation intensity analysis is performed on the multi-channel feature maps corresponding to each convolution and downsampling layer, and the execution of convolution and downsampling operations is judged.

[0037] The multi-channel feature maps output by the parallel fusion network composed of convolution and attention mechanisms and the sparse connection U-shaped network are fused to determine the arrival candidate regions of P-waves and S-waves, forming the arrival distribution curves of P-waves and S-waves.

[0038] In this embodiment, the collected DAS data is first segmented and resampled. This embodiment selects to segment and resample the DAS data into 60-second segments, followed by adaptive morphological filtering to obtain denoised DAS data. After obtaining the denoised DAS data, it is input into a parallel fusion module of convolution and attention mechanisms and a sparsely connected U-shaped attention module, respectively. The parallel fusion module of convolution and attention mechanisms, through parallel convolution operations and self-attention mechanisms, simultaneously focuses on both local and global features, which is more consistent with the physical characteristics of DAS data than traditional convolution and self-attention mechanisms, and concentrates the model's attention on the locations where P-waves and S-waves are concentrated. Then, the sparsely connected U-shaped attention module is combined to downsample and extract features, and only one skip connection is made in the entire U-shaped structure, ensuring that the model pays attention to the distribution trends of P-waves and S-waves while maintaining low computational cost. The depth features obtained through the above steps, by dividing the DAS data phase picking task into two tasks, ensure low computational cost in the feature extraction process while ensuring that the depth features simultaneously contain both global and local information of the original DAS data. Finally, the depth features obtained in the above steps are used to pick up seismic phases through depthwise separable convolutions and linear layers, thus completing the lightweight DAS seismic phase picking network based on sparse connections.

[0039] Adaptive morphological filtering specifically involves: firstly, generating dynamic structuring elements; and secondly, based on the local signal variance in the time dimension. (The degree of local signal fluctuation at the current moment reflects the noise intensity) Adjust the length of the structural element. (The length of the dynamic structural element in the time dimension). In the spatial dimension, based on the correlation between adjacent channels. Use this to select the shape of the structural element.

[0040] ;

[0041] ;

[0042] in, Sampling rate, It is based on the noise threshold stored in the database, which is set through dataset statistics. The [x] function represents the rounding function, making the length of the structuring element an integer number of points. ρ k，k+1 Let represent the Pearson correlation coefficient between the k-th channel and the (k+1)-th channel in the multi-channel feature map over the time dimension. Finally, the filtered signal is obtained through an erosion-dilation cascade.

[0043] ;

[0044] in, For opening operation, For closing operations, The shape of the spatiotemporal structural element, that is, the morphological structural element. The input is the raw DAS data. For DAS data after adaptive morphological filtering, the opening operation is performed. AND closing operation The definitions are as follows: ; ;in Indicates corrosion. The opening operation removes positive spikes smaller than the structuring element from the signal while preserving a larger smooth structure; it is often used to suppress isolated positive noise. The closing operation fills in negative depressions smaller than the structuring element in the signal, enhancing continuity and connectivity; it is often used to suppress fine negative noise or connect weak stripes across channels. t represents the index in the time dimension of the DAS data. , This represents the number of sampling points in the time dimension of the DAS data.

[0045] It should be noted that, for example, if the fiber optic cable is several kilometers long and the spatial sampling interval is 1 meter, then the fiber optic cable will have thousands of channels. Each channel represents the vibration response at a fixed location on the fiber. During the acquisition phase, the spatial location information, burial depth, and nearby structural parameters of each channel are recorded simultaneously. These are spatial metadata used to reflect the geological background and laying status of each channel. For the time series output by each channel, it is divided into multiple time windows along the time axis. Each window corresponds to a continuous time segment, such as 10 seconds, 30 seconds, or 60 seconds. The time segment is dynamically set according to the signal characteristics. The speed and complexity of the signal change are judged based on the average wave velocity range, event frequency bandwidth, and amplitude envelope length of the current segment: if the wave velocity is high, the frequency bandwidth is wide, or the envelope is short, it indicates that the signal changes rapidly and the event is concentrated, so a short window is used to capture transient features; if the wave velocity is low, the frequency bandwidth is narrow, or the envelope is long, a long window is used to ensure the integrity of the seismic phase morphology.

[0046] It should be noted that the specific process of segmenting the DAS data of the target area to form the DAS signal sequence of the target area is as follows: First, the raw continuous DAS signal stream from the fiber optic distributed acoustic wave sensing system is received. This signal contains vibration response data of multiple spatial sampling points in the target area along the time dimension. Based on the spatial distribution information of the optical cable and the sampling frequency parameters recorded during acquisition, the raw data is organized into a multi-channel continuous signal in chronological order. Then, according to the duration of each segmentation window of the target area DAS data, the target area DAS data is divided along the time axis in a sliding or non-overlapping manner to generate multiple continuous segments. Each segment contains the signal response of all channels within a fixed time length. Finally, all segments are arranged sequentially according to the acquisition order to form the target area DAS signal sequence. This signal sequence maintains complete coverage in time and corresponds to the continuous distribution of the optical cable in space, providing a spatiotemporally consistent input basis for subsequent local window segmentation, adaptive filtering, and multi-scale feature extraction.

[0047] like Figure 2 As shown, Figure 2 The flowchart of the lightweight DAS phase picking method based on sparse connections provided in this embodiment of the invention first segments the DAS data of the target area and performs scale adjustment and morphological processing on a local time window to generate an enhanced signal feature sequence. Then, multi-scale feature extraction and downsampling are performed through a sparsely connected U-shaped network encoding path, and the comprehensive complexity index value of each downsampling layer is calculated. The next operation is determined by judging whether the complexity reaches a preset threshold. If the complexity is high, downsampling is continued to extract multi-scale features; otherwise, the convolution branch is entered to fuse with the U-shaped network features to remove low-response channels. Finally, the fused feature sequence is processed to generate three-dimensional distribution curves of P-wave and S-wave arrival, realizing phase identification and spatial distribution visualization of DAS signals.

[0048] like Figure 3 As shown, Figure 3The overall flowchart of the end-to-end phase picking model for DAS data provided in this embodiment of the invention begins with the preprocessing of DAS data (segmentation, resampling, adaptive morphological filtering), followed by feature extraction through a core parallel dual-branch structure: one branch (convolutional branch) focuses on capturing the local concentrated location features at the arrival of P / S waves, while the other branch (attention branch) focuses on analyzing the global distribution trend at the arrival of P / S waves; the features from these two branches are integrated through a parallel fusion module of convolution and attention, and their output is further optimized by a sparsely connected U-shaped attention module, finally generating accurate phase picking results through a linear layer. In short, this model achieves high-precision and automated identification of phases in DAS signals by fusing local and global information. The preprocessing of DAS data includes segmenting the target region DAS data to form a target region DAS signal sequence. Specifically, the target region DAS data is segmented and analyzed to obtain first segmentation windows. Temporary vibration response data along the optical fiber within each first segmentation window is acquired, including acoustic energy intensity, instantaneous rate of change of signal, and signal waveform variance. The acoustic energy intensity, instantaneous rate of change of signal, and signal waveform variance are fused to obtain a scale adjustment factor for each first segmentation window. This scale adjustment factor is used to evaluate the activity level of the signal within each first segmentation window. The duration of each segmentation window is obtained based on the scale adjustment factor. The first segmentation windows are then adjusted according to their durations, thereby segmenting the target region DAS data to form a target region DAS signal sequence.

[0049] It should be noted that the specific process of fusing acoustic energy intensity, signal instantaneous change rate, and signal waveform variance to obtain the scale adjustment factors for each first segmentation window of the target region DAS data is as follows: Within each first segmentation window of the target region DAS data, acoustic energy intensity, signal instantaneous change rate, and signal waveform variance are first extracted. Acoustic energy intensity is obtained by averaging the squared signal amplitudes within the window, reflecting the overall energy level of vibration during that time period. The signal instantaneous change rate is obtained by calculating the amplitude of every two consecutive sampling points within the window and dividing it by the sampling interval, forming the instantaneous change rate sequence for that window. The mean of this sequence is then calculated to obtain the average instantaneous change rate for that window, used to quantify the overall level of rapid signal change during that time period. The waveform variance is obtained by acquiring the amplitude sequence of all sampling points within the window, calculating the squared deviation of this amplitude sequence from its mean, and averaging the results. Each feature is first normalized, mapping the feature values ​​to the range of 0 to 1 to eliminate the influence of different dimensions and amplitude differences on subsequent fusion. Subsequently, a nonlinear mapping, such as an exponential or square function, is applied to each normalized feature to further amplify high-value features and attenuate low-value features. The three nonlinearly mapped values ​​are then multiplied element-wise to form the fusion value of the window, which serves as the scaling factor for the first segmentation window. This factor simultaneously reflects the nonlinear interaction effects between energy level, rate of change, and waveform complexity, making it suitable for adaptive signal segmentation and filtering. When any feature is low, the product is significantly suppressed, preventing the window from being misjudged as a high-activity region. Conversely, when all three features are at high levels, the product is synchronously amplified, highlighting key vibration segments with strong energy, high rate of change, and significant waveform fluctuations. This effectively improves the sensitivity of window segmentation to high-value vibration regions while suppressing false triggers caused by noise or single-indicator anomalies, making the adaptive window adjustment more robust and accurate. If a feature is zero or abnormally low, the product will directly become zero, potentially losing some signal. Therefore, the minimum value of the three elements is limited to avoid completely erasing window information. Finally, the duration of the first segmentation window is dynamically adjusted based on its scale adjustment factor to preserve rapid amplitude abrupt changes and high-frequency details. Windows with smaller scale adjustment factors can have their duration extended or their structuring elements enlarged to smooth background signals and reduce redundant information. This nonlinear fusion method allows for precise control of segmentation and filtering parameters based on different window characteristics, achieving high-resolution, adaptive processing of DAS signals across the entire time and spatial domains. This dynamically generates the duration of each segmentation window within the entire target region's DAS data, enabling adaptive processing of regions with different signal characteristics.

[0050] It should be noted that the specific process for obtaining the duration of each segmentation window of the target region DAS data is as follows: The scale adjustment factor of each first segmentation window of the target region DAS data is mapped to a window length mapping table stored in the database. This database records the optimal time length or structuring element scale corresponding to different factor values. For example, the scale factor is divided into several intervals, each interval corresponding to a recommended segmentation window length. The database is searched based on the scale adjustment factor of each first segmentation window to obtain the target segmentation length for each window. Then, the system performs dynamic adjustment operations on each window: if the target length is shorter than the original length, the original window is further subdivided into several local sub-windows; if the target length is longer than the original length, several consecutive low-energy, stable windows can be merged into a longer window to reduce redundant segmentation and smooth signal changes. This mapping method ensures that windows with high energy, rapidly changing waveforms, and high local complexity are assigned shorter segmentation lengths to capture details, while windows with low energy and stable waveforms are assigned longer segmentation lengths to reduce redundant computation. Finally, after mapping adjustment, the time length of each first segmentation window is optimized to an adaptive scale suitable for subsequent filtering and feature extraction, realizing dynamic and segmented processing of the entire DAS signal sequence.

[0051] The morphological structural elements are dynamically adjusted based on each local time window. The specific process is as follows: the DAS data of the target region is segmented to form the DAS signal sequence of the target region. It is divided into several local segments according to time to obtain each local time window. The temporal vibration response data along the optical fiber of the DAS data of the target region in each local time window is extracted to obtain the scale adjustment factor of each second local time window of the DAS data of the target region. The scale adjustment factor of each second local time window of the DAS data of the target region is used to evaluate the vibration mode stability and energy distribution change characteristics of the local signal at a fine-grained time scale. The morphological structural elements are dynamically adjusted according to the scale adjustment factor of each second local time window of the DAS data of the target region.

[0052] It should be noted that the specific process for obtaining each local time window is as follows: First, the DAS data of the target area is initially segmented to form several continuous signal segments. Each segment retains information from the original sampling sequence, including time-series vibration response and amplitude data. These segments constitute the DAS signal sequence of the target area. The length of each segment can be determined based on the preliminary analysis results, acquisition conditions, and event characteristics, but at this stage, it still belongs to a relatively large time scale segmentation. Subsequently, each signal sequence segment is further refined into several local time windows, each containing a continuous sampling point data. Each local time window contains a complete signal amplitude sequence and time series information. Features such as acoustic energy intensity, instantaneous rate of change of the signal, and waveform variance can then be calculated for nonlinear fusion to generate a scale adjustment factor. This enables adaptive window length and filtering parameter settings, allowing for refined processing of local seismic phase characteristics while maintaining global continuity, achieving multi-scale, adaptive analysis of DAS data in the time dimension.

[0053] It should be noted that the temporal vibration response data along the fiber optic cable of the DAS data in the target region within each local time window does not refer to newly introduced data, but rather to the temporal vibration response data along the fiber optic cable of the DAS data in the target region within each local time window. However, when entering the scale factor calculation stage, this data is further refined from the initial coarser time segments into local windows for analysis. Therefore, it naturally contains the same type of information as the first segmentation window, namely the temporal vibration response data along the fiber optic cable of the DAS data in the target region within each local time window, including local statistical features derived from the original vibration sequence such as energy intensity, instantaneous rate of change, and waveform variance. The first segmentation window provides the fundamental vibration structure over a longer time scale, while each local time window is actually a sub-window obtained by further refining this fundamental segmentation. They belong to the internal time sub-intervals of the first segmentation window, so physically, they do not introduce new data sources; they only change the time scale and granularity, allowing the same vibration response sequence to be recalculated and formatted within a shorter time range. Based on this hierarchical relationship, the temporal vibration response extracted from each local time window is used to form the input content of the scale adjustment factor of the second local time window. In essence, it is a localized and fine-grained reanalysis of the existing vibration data in the first segmentation window, so that the scale adjustment factor can truly reflect the stability and energy change trend of the vibration mode at a small time scale, and is ultimately used to dynamically update the structural element.

[0054] It should be noted that the specific process of dynamically adjusting the morphological structuring elements based on the scale adjustment factors of each second local time window of the target region DAS data is as follows: First, the scale adjustment factors of each second local time window are mapped to the structuring element parameter mapping table stored in the database. This database records the optimal structuring element time scale and geometric morphological parameters corresponding to different scale factor values. The scale factor is divided into several numerical intervals, and each interval corresponds to a set of structuring element length parameters and morphological description parameters. The mapping table is then searched based on the scale adjustment factor of each second local time window. The target structuring element parameters and target sampling parameters are obtained from the mapping database. The structuring element parameters include the time scale and geometric morphology, and the sampling parameters include the sampling rate or the number of sampling points. The target structuring element parameters returned by the database are directly overwritten with the original window configuration, allowing the window to use the new structuring elements in subsequent morphological filtering. Subsequently, the target sampling rate is compared with the current sampling rate: if the target sampling rate is higher than the current sampling rate, interpolation resampling is performed on the window signal; if the target sampling rate is lower than the current sampling rate, low-pass filtering is first performed (filtering out high-frequency signals that exceed the frequency range that can be represented by the sampling frequency stored in the CNC library for resampling the signal), and then sampling points are extracted according to the target sampling rate. If the target sampling rate is equal to the current sampling rate, it remains unchanged, thereby ensuring that the signal can be directly used for filtering operations in the adjusted sampling state.

[0055] like Figure 4 As shown, Figure 4 This diagram illustrates the parallel convolution-attention feature fusion module, employing a dual-path parallel structure: the upper path extracts local features through 1x1 convolutions and 3x3 depthwise separable convolutions; the lower path utilizes a multi-head attention mechanism to generate queries, keys, and values, which are then fused with the convolution path outputs through element-wise addition and multiplication after the attention weights are calculated using the Softmax activation function. This design aims to simultaneously capture spatial details and long-range dependencies, where Conv represents convolution, DWConv represents depthwise separable convolution, Q / K / V represents query / key / value, Softmax represents the normalized exponential function, ⊕ represents element-wise addition, and ⊗ represents element-wise multiplication.

[0056] The upper path extracts local features through 1x1 convolution and 3x3 depthwise separable convolution, including accurate characterization and identification of vibration patterns in the target region. The specific process is as follows:

[0057] The DAS-enhanced signals of the target region obtained after morphological structuring are arranged in their original temporal order to form a continuous enhanced signal feature sequence. The enhanced signal feature sequence is input into a convolution branch, in which several groups of convolution kernels of different lengths are set, each convolution kernel corresponding to a specific time scale. Each convolution kernel is slid along the time axis on the sequence, and at each sliding position, a signal segment of the current position and its neighborhood is taken. The segment is multiplied element-wise with the convolution kernel weights and summed to obtain the local feature response value of each convolution kernel at the current position. The local feature response value of each convolution kernel at the current position is used to evaluate the degree of matching between the local features of the DAS signal of the target region and the convolution kernel weight pattern at that time point and its neighborhood.

[0058] Local energy and waveform change data of the signal are obtained in the signal range covered by each convolution kernel when it slides to a certain time point, including the average value of the square of the DAS signal amplitude in the target region and the instantaneous rate of change of the signal.

[0059] If the average value of the squared amplitude of the DAS signal in the target region and the instantaneous rate of change of the signal both exceed the corresponding range of the squared amplitude and the range of the instantaneous rate of change of the signal stored in the database, then the local feature response value of the convolution kernel at the current position corresponding to the signal range covered by the convolution kernel when it slides to a certain time point is corrected; otherwise, it is not necessary to correct the local feature response value of the convolution kernel at the current position.

[0060] In this embodiment, the focus of the parallel fusion module of convolution and attention mechanisms is to simultaneously apply convolution operations and a self-attention module to the input features. In the convolution operation, the input features are first remapped using a 1x1 convolution kernel:

[0061] X1 = X × W1x1 + b1x1; where X represents the DAS feature map output from a previous convolution / feature extraction layer after being input into the convolutional self-attention structure, and W1x1 and b1x1 are the learnable parameters of the 1x1 convolution.

[0062] Then, depthwise separable convolution with a 3x3 kernel is used to extract local texture features from the input features, i.e., depthwise convolution and pointwise convolution are performed sequentially:

[0063] Y = X1 × Wd × Wp + bp; where Wd is the learnable parameter of depthwise convolution in depthwise separable convolution, and Wp and bp are the learnable parameters of pointwise convolution.

[0064] In the self-attention module, the input DAS data X is first remapped using a 1x1 convolution:

[0065] X'=X×W1x1; where X is the channel remapping result obtained by 1×1 convolution of the input feature map X, which is used to construct the basic feature representations of Q, K, and V.

[0066] Then, the query (Q), key (K), and value (V) are obtained through three different 3x3 convolutional kernels:

[0067] Q=X'×W q 3x3 ;

[0068] K=X'×W k 3x3 ;

[0069] V=X'×W v 3x3 ;

[0070] Among them W q 3x3 W k 3x3 With W v 3x3 These are the learnable parameters for obtaining three different 3x3 convolutions of Q, K, and V, respectively. For convolution computation, attention maps are calculated using Q and K and then applied to V to obtain the final feature map:

[0071] ;

[0072] ;

[0073] ;

[0074] in It is the dimension of the key, used for scaling. Indicates to Take the exponent for each element. This represents the summation of the exponents of all elements. For matrix multiplication, Represents the matrix The matrix is ​​transposed, the Attention Map represents the relevance weights, the Softmax() function normalizes the row dimensions of the matrix, and Z is the attention-weighted feature output. These are further integrated and remapped using a 1×1 convolution. Finally, the remapped feature map is combined with the initial input DAS data. Multiply to obtain the final output: Y' = X⊙Z

[0075] At the end of the parallel fusion module of convolution and attention mechanisms, the convolution operation is added point by point to the output of the self-attention module to obtain the final module output.

[0076] It's important to note that the parallel fusion network of convolution and attention mechanisms is a signal processing framework that simultaneously incorporates convolution operations and attention mechanisms. The convolution branch primarily extracts signal features at local time scales, capturing rapid amplitude changes and local energy variations within a short timeframe. The attention branch, on the other hand, calculates the dependencies and global correlations between time points throughout the entire time series. Through dynamic weights, it emphasizes segments with high energy and rapid phase changes over short periods, highlighting key seismic phase segments in the sequence. These two branches operate in parallel, meaning the signal is simultaneously fed into both branches for parallel computation. This ensures that both subtle local waveform features are captured and global temporal dependencies are considered. In the convolution branch, multiple convolutional kernels of varying lengths slide across the time series. Each kernel performs continuous convolution operations along the time dimension, weighting and summing the signals at each time point and its neighborhood to form a local feature response. The feature map output by the convolution reflects the strength of local energy changes and waveform variations along the time axis, providing initial candidate regions for concentrated seismic phase locations.

[0077] It's important to note that the specific time scale corresponding to each convolutional kernel refers to the different lengths of time each kernel can sense during its design; in other words, its receptive field covers a different temporal range, thus enabling it to capture signal features across different time spans. For example, a pre-defined short-time-scale convolutional kernel is shorter, and when sliding across the sequence, it focuses on a short signal segment at a time, capturing detailed features such as instantaneous amplitude changes and rapid waveform variations. Conversely, a long-time-scale convolutional kernel is longer, focusing on a longer signal segment at a time, and can extract local energy change trends, low-frequency fluctuations, or slower vibrational patterns. Therefore, the time scale of each convolutional kernel actually refers to the length of time the kernel operates on the time series, determining the speed and span of signal changes it can detect. Combinations of different convolutional kernels create multi-time-scale feature extraction capabilities, capturing both short-term details and long-term trends.

[0078] It's important to note that during convolution calculations, the convolution kernel slides point-by-point along the time axis of the enhanced signal feature sequence. At each position, the system extracts a continuous signal segment consisting of that time point and several surrounding sample points. The length of this segment is the same as the length of the convolution kernel, which is the receptive field of the kernel. Then, each sample value in this signal segment is multiplied one-to-one with the corresponding weight in the convolution kernel. That is, the first sample value of the signal segment is multiplied by the first weight of the convolution kernel, the second sample value by the second weight, and so on until the end of the segment. After element-by-element multiplication, all products are summed in chronological order to obtain a scalar value. This scalar value is the response value of the convolution kernel at the current position, reflecting the degree of matching between the convolution kernel weight pattern and the signal segment. The convolution kernel is then slid along the sequence, and the same operation is repeated each time. The corresponding response value is calculated for each time point and its neighborhood segment in the sequence. Finally, a local feature response sequence of the convolution kernel on the entire enhanced signal sequence is formed. This sequence can characterize the degree of fit between the signal and the weight pattern of the convolution kernel at different time points, thereby capturing the characteristics of local energy abrupt changes and waveform changes.

[0079] It should be noted that the local feature response value is the mathematical quantification output of the convolution kernel on the signal features of its covered region, while the local energy and waveform change data are physical quantities that describe the signal properties of that region.

[0080] It should be noted that the amplitude change of adjacent sampling points can be expressed by the instantaneous rate of change of the signal, which is the ratio of the difference in amplitude between two consecutive sampling points to the sampling interval. This reflects the rapid degree of change of the signal over time. The average of the squared amplitudes of the DAS signal in the target region is used to quantify the local energy. That is, within a selected time interval, the average of the squared amplitudes of all sampling points in that interval is taken to obtain the energy level of that interval. The specific method is as follows: first, the enhanced DAS signal is extracted within the required time interval, the amplitude of each sampling point is taken and squared, then these squared values ​​are summed and divided by the total number of sampling points to obtain the average squared amplitude of the interval, which is the local energy. For the instantaneous rate of change, within the same time interval, for every two adjacent sampling points, the amplitude of the latter sampling point is subtracted from the amplitude of the former sampling point, and then divided by the sampling interval to obtain the amplitude change rate at that moment. By repeating the calculation for all adjacent sampling points throughout the entire time interval, a complete sequence of instantaneous rate of change can be obtained. This sequence reflects the strength and speed of the waveform change of the signal within that interval.

[0081] It should be noted that when each convolution kernel slides to a certain time point, the signal range covered by the convolution kernel is first obtained, that is, the target region DAS signal sequence of the current position and its neighborhood. The local energy and waveform changes of the range are calculated, including the average value of the square of the signal amplitude and the instantaneous rate of change. Then, this signal range is multiplied element-wise with the weight of the convolution kernel and summed to obtain the local feature response value of the convolution kernel at the current position.

[0082] It should be noted that the specific process for correcting the local feature response value of the convolutional kernel at a certain time point, corresponding to the signal range covered by the kernel, is as follows: If the average square of the DAS signal amplitude and the instantaneous rate of change of the signal both exceed the corresponding ranges of the square of the signal amplitude and the range of the instantaneous rate of change stored in the database, the system first looks up the convolutional response correction ratio corresponding to the average square of the amplitude and the instantaneous rate of change in the mapping table stored in the database. This mapping table strictly corresponds different signal feature value ranges with the adjustment ratio of the local feature response value of the convolutional kernel at the current position, ensuring that the correction amount matches the signal features. For ranges exceeding the upper limit, the system linearly adjusts the local feature response value of the convolutional kernel at the current position according to the amplification ratio provided by the mapping table, so that the response value is consistent with the actual local signal feature intensity, thereby enhancing the response of high-energy segments with rapidly changing waveforms; for ranges below the lower limit, the system looks up the corresponding reduction ratio in the mapping table and reduces the local feature response value of the convolutional kernel at the current position accordingly to suppress the response of low-energy or stable waveform segments. During the adjustment process, the system strictly follows the proportions provided in the mapping table to perform multiplication corrections.

[0083] like Figure 5 As shown. Figure 5 This diagram illustrates the architecture of a sparse attention module based on U-Net and skip connections. Its core is a U-shaped network with an encoder-decoder structure. Input features are downsampled through multiple 3x3 convolutions, layer normalization, and 3x3 depthwise separable convolutions. The right branch then calculates Q / K / V through an attention mechanism and establishes a cross-layer connection with the upsampling path. The module integrates features from both shallow and deep layers through skip connections, ultimately outputting features via 1x1 convolutions and upsampling. This structure reduces redundant computation through sparse connections, emphasizes the key feature transfer using Softmax weighting, and improves feature reconstruction efficiency. Here, Conv represents convolution, DWConv represents depthwise separable convolution, Layer Norm represents layer normalization, Upsample represents upsampling, Q / K / V represents query / key / value, Softmax represents the normalization exponential function, ⊕ represents element-wise addition, and ⊗ represents element-wise multiplication. The integration of features from both shallow and deep layers through skip connections includes channel activation intensity analysis of the multi-channel feature maps corresponding to each convolution and downsampling layer. The specific process is as follows:

[0084] A continuous sequence of enhanced signal features is input into the convolutional branch of a sparsely connected U-shaped network. Convolution and downsampling operations are performed in the encoding path of the U-shaped network to extract the spatiotemporal feature responses at each scale layer by layer to generate a multi-feature response matrix. The corresponding multi-channel feature map is automatically constructed in each downsampling layer and denoted as the multi-channel feature map corresponding to each downsampling layer.

[0085] Extract the feature map automatically constructed for each downsampling layer, and take the amplitude of the feature vector of each channel in the feature map automatically constructed for each downsampling layer point by point in the time dimension. Then, sum the amplitudes and divide them by the total number of feature vectors in the channel to obtain the average response value of each channel corresponding to each downsampling layer.

[0086] The average response value of each channel corresponding to each downsampling layer is compared with the average response threshold of the channel corresponding to the downsampling layer stored in the database. If the average response value of a channel corresponding to a downsampling layer is lower than the average response threshold of the channel corresponding to the downsampling layer, the channel corresponding to the downsampling layer is marked as a low response channel and its characteristics are not retained in the skip connection. Otherwise, it is not necessary to mark the channel corresponding to the downsampling layer as a low response channel.

[0087] In this embodiment, the continuous enhanced signal feature sequence is input into the convolutional branch of the sparsely connected U-shaped network, which involves inputting DAS data through a 1x1 depthwise separable convolution followed by layer normalization.

[0088] X norm =LayerNorm((X*Wd)*Wp+bp);

[0089] Subsequently, two 3x3 depthwise separable convolutions are performed to achieve dimensionality reduction, and then a low-resolution feature map is obtained through the same self-attention module as in the parallel fusion module of convolution and attention mechanisms. The layer normalization operation LayerNorm is performed on the channel dimension:

[0090] ;

[0091] Where b, c, t, and n represent the four dimensions of data batch size, number of channels, time duration, and spatial duration, respectively. μ b,t,n With σ 2 b,t,n ε and γ represent the mean and variance of the data along the channel dimension, respectively, where ϵ is a very small constant used to prevent division by zero. c ,β c These are two learnable parameters used for scaling and offset.

[0092] Then, two 3x3 depthwise separable convolutions are performed to reduce dimensionality. Next, a low-resolution feature map, Xattention, is obtained through the same self-attention module as in the parallel fusion module of convolution and attention mechanisms. Finally, two upsampling operations are performed to restore the low-resolution feature map to the original input size.

[0093] Where Xattention_up represents the upsampled attention feature map, which will then be... The final module output is obtained by performing a single jump connection in a U-shaped structure with the input DAS data. Two upsampling operations are performed. All methods use bilinear interpolation, and each upsampling restores the feature scale to the same feature scale as the corresponding coding layer features.

[0094] It's important to note that skip connections are a neural network architecture design technique. They refer to the direct transfer of features computed in the encoding phase (downsampling or feature extraction layer) to the corresponding decoding phase (upsampling or feature recovery layer), rather than simply propagating through layers of the network. In this embodiment, the convolutional branch of the sparsely connected U-shaped network performs only one skip connection, maintaining low computational cost while ensuring the model pays attention to the distribution trends of P-waves and S-waves. Specifically, in a U-shaped network or U-Net structure, the signal undergoes convolution, pooling, or downsampling operations in the downsampling phase to extract multi-scale features. These features lose some high-resolution local information during layer-by-layer compression. The role of skip connections is to directly transfer these high-resolution encoded features to the upsampling decoding phase and fuse them with the current features in the decoding layer (e.g., concatenation or addition), thereby integrating local details with global semantic information.

[0095] Finally, the outputs of the two modules are multiplied point by point in the model, and then the data is processed through depthwise separable convolution and linear layers to finally complete the phase picking of the DAS data.

[0096] The determination of whether to perform convolution and downsampling operations is as follows: statistical analysis is performed on the multi-channel feature maps corresponding to each downsampling layer in both the time and channel dimensions to obtain the global statistics of each downsampling layer, including channel mean, channel variance, time energy variance, and time-frequency energy entropy. The channel mean, channel variance, time energy variance, and time-frequency energy entropy are normalized and fused to obtain the comprehensive complexity index value of each downsampling layer. The comprehensive complexity index value of each downsampling layer is used to reflect the complexity and information density of the features of each downsampling layer.

[0097] The overall complexity index value of each downsampling layer is compared with the overall complexity index threshold of the downsampling layer stored in the database. If the overall complexity index value of a downsampling layer is lower than the overall complexity index threshold of the downsampling layer, further convolution and downsampling operations are automatically stopped. If the overall complexity index value of a downsampling layer is higher than or equal to the overall complexity index threshold of the downsampling layer, further convolution and downsampling operations are continued.

[0098] It should be noted that the specific process for obtaining the comprehensive complexity index value of each downsampling layer is as follows: Within each downsampling layer, the system first extracts the corresponding temporal energy variance and time-frequency energy entropy. The temporal energy variance quantifies the energy fluctuation of the enhanced signal feature sequence in the time dimension, while the time-frequency energy entropy characterizes the uncertainty of the signal energy in the frequency distribution. To ensure the comparability of each index during fusion, each index is first subjected to min-max normalization, linearly mapping its value to the interval between 0 and 1, allowing features at different scales to participate in subsequent calculations under the same standard. Subsequently, the normalized indices undergo nonlinear transformation. The channel mean and temporal energy variance are compressed using a square root method to suppress the influence of extreme values ​​on the comprehensive complexity calculation, while the channel variance and time-frequency energy entropy are amplified by logarithmic mapping to enhance the contribution of weak energy fluctuations to the complexity assessment. After the transformation, the system adaptively fuses the four normalized parameters within the current downsampling layer. Specifically, it determines the implicit weight by calculating the correlation between each parameter and the comprehensive complexity index value of the previous layer. The correlation is quantified using Pearson correlation coefficient or cosine similarity to measure the synchronicity and consistency of the current index with the complexity change of the previous layer. The calculated implicit weights are then assigned to each parameter. Finally, the system combines the normalized indices using a weighted summation to obtain the comprehensive complexity value of the current downsampling layer. This value reflects the feature complexity, local vibration modes, and energy distribution of the current layer signal, and provides a basis for deciding whether to continue downsampling and for feature channel selection. Throughout the downsampling encoding process, each layer calculates the comprehensive complexity index using this method, and the results are used to dynamically control the downsampling strategy and feature preservation strategy. This ensures that important vibration modes within the stable low-frequency characteristic energy range, especially the energy transition from P-waves and S-waves to the time region, are fully preserved. This ensures that the model accurately represents the vibration modes of the target region and captures them at multiple scales while maintaining low computational cost.

[0099] It should be noted that the feature maps automatically constructed by each downsampling layer are used to describe the distribution of signal energy and temporal structure at different scales, providing multi-level feature support for subsequent attention modulation and decoding reconstruction.

[0100] The multi-channel feature maps output by the parallel fusion network composed of convolution and attention mechanisms and the sparse connection U-shaped network are fused. Specifically, the multi-channel feature maps output by the parallel fusion network of convolution and attention mechanisms are arranged according to the time dimension and the channel dimension to obtain the local vibration mode response sequence. At the same time, the multi-channel feature maps output by the sparse connection U-shaped network are arranged according to the same time and channel dimensions to obtain the multi-scale global vibration mode sequence.

[0101] The local vibration mode response sequence is matched point-by-point with the multi-scale global vibration mode sequence, and then linearly superimposed to obtain the fused multi-channel feature sequence.

[0102] It should be noted that the specific process of point-by-point matching and linear superposition of the local vibration mode response sequence and the multi-scale global vibration mode sequence is as follows: First, the local vibration mode response sequence and the multi-scale global vibration mode sequence are aligned according to the time dimension, so that the channel features of the two sequences correspond one-to-one at the same time point. For each time point, the amplitude feature of each channel in the local vibration mode response sequence is linearly weighted with the amplitude feature of the same channel in the corresponding multi-scale global vibration mode sequence. The weight is calculated by the proportion of the amplitude of the channel in the local sequence to the sum of the amplitudes of all channels at that time point, obtaining the local feature weight. At the same time, the proportion of the amplitude of the channel in the global sequence to the sum of the amplitudes of all channels at the same time point is calculated, obtaining the global feature weight. These two weights are normalized so that their sum is 1. Then, the local feature value and the global feature value are weighted and summed according to the normalized weights to obtain the fused channel amplitude. By repeating this process for all channels at each time point in the time series, a fused multi-channel feature sequence that is completely corresponding in both time and channel dimensions is formed. This ensures that local energy mutation information is preserved while reflecting the correlation of global vibration modes during the fusion process. After weighting, the linear superposition results of each channel at each time point are combined into a new multi-channel feature vector. All time points are then processed sequentially along the time series to obtain a fused multi-channel feature sequence that is completely corresponding in both time and channel dimensions. Subsequently, the fused multi-channel feature sequence is normalized to adjust the amplitude of each channel to a uniform scale, and smoothing filtering is performed when necessary to eliminate local noise interference. This ensures that the fused sequence reflects both local amplitude peak changes and global vibration mode characteristics, providing accurate input data for subsequent determination of P-wave and S-wave arrival candidate regions.

[0103] The process of determining the arrival candidate regions of P-waves and S-waves is as follows: the fused multi-channel feature sequence is scanned point by point along the time axis to obtain the amplitude of each channel at each time point.

[0104] A preset monitoring period is defined. Within each monitoring period, the local feature response values ​​of each convolutional kernel at its current position are extracted. These local feature response values ​​are then plotted on a time axis to form a local feature response sequence. The average amplitude of this sequence is calculated to obtain the global amplitude for each monitoring period. The global amplitude for each monitoring period is compared with pre-stored typical amplitude thresholds for P-waves and S-waves in a database. The frequency range of the local feature response sequence is determined, and then jointly labeled with the fiber optic channel to form candidate P-wave and S-wave arrival areas containing both time interval and channel information.

[0105] It should be noted that the global amplitude of each monitoring time period is compared with the pre-stored typical amplitude thresholds for P-waves and S-waves in the database, and the frequency range of the local characteristic response sequence is determined. Specifically, for P-waves, it is necessary to determine whether the signal amplitude during that time period is lower than the typical P-wave amplitude threshold, and whether its frequency falls within the typical frequency range of P-waves. For S-waves, it is determined whether the amplitude is higher than the S-wave amplitude threshold and whether its frequency is within the typical frequency range of S-waves. Through this joint determination of amplitude and frequency, the existence of different wave types within the same time period can be distinguished, while avoiding misclassification of non-P-wave or S-wave background signals as candidate phases, thereby accurately marking the possible arrival areas of P-waves or S-waves in each channel both spatially and temporally.

[0106] It should be noted that, due to the different amplitude characteristics of P-waves and S-waves in vibration signals, thresholds are generally set separately for P-waves and S-waves to ensure that their respective distinct vibration signals can be detected. Joint labeling refers to labeling all fiber optic channels that simultaneously meet the amplitude and frequency conditions within a given time period as candidate regions, forming a candidate P-wave or S-wave arrival area that includes information from all relevant channels within that time period. This joint labeling method integrates information from multiple channels that respond simultaneously in space, reducing false positives from single-channel noise while preserving the spatial structural characteristics of multi-channel vibrations, providing a reliable basis for subsequent phase arrival location.

[0107] The process of forming the arrival distribution curves of P-waves and S-waves is as follows: the time periods marked as candidate arrival intervals of P-waves and S-waves within each monitoring time period and their corresponding amplitude information are arranged in chronological order. Combined with the spatial distribution of each optical fiber channel, a channel amplitude matrix is ​​constructed. The amplitude is normalized and mapped on the time axis according to the strength of the amplitude, and the area with higher amplitude is highlighted to form the three-dimensional distribution curves of arrival of P-waves and S-waves.

[0108] It should be noted that in the specific process of generating the arrival distribution curves of P-waves and S-waves, it is first necessary to organize the time periods marked as candidate P-wave and S-wave arrival intervals and their corresponding amplitude information within each monitoring time period of the DAS signal in the target area. The system arranges all candidate intervals in chronological order according to the start and end times of each monitoring time period, ensuring that the amplitude data of each time period is continuous and orderly in the time dimension. For each time point, the signal amplitude information of the corresponding fiber optic channel is extracted, and these amplitudes are correlated with the actual spatial layout of the fiber optic channel, thereby constructing a two-dimensional channel amplitude matrix. The rows represent the spatial sequence position of the fiber optic channel, and the columns represent the sampling points or monitoring time periods in the time dimension. This channel amplitude matrix not only reflects the amplitude magnitude of each channel at each time point but also retains the spatial distribution information of the vibration signal by the fiber optic array, providing a foundation for subsequent spatial-temporal characteristic analysis of P-waves and S-waves. After constructing the channel amplitude matrix, the amplitude data in the matrix is ​​normalized, mapping each amplitude to a uniform standard range, such as 0 to 1, according to its strength. This eliminates the difference in absolute amplitude magnitude between different channels, ensuring fair comparison of local energy peaks in subsequent processing. A linear mapping method can be used during normalization, mapping the minimum amplitude to 0, the maximum amplitude to 1, and other amplitudes to the range of 0 to 1 based on their relative magnitude. After normalization, regions with higher amplitudes are clearly highlighted during visualization, effectively reflecting the energy concentration characteristics of P-waves and S-waves. Subsequently, the normalized channel amplitude matrix undergoes further smoothing to reduce amplitude fluctuations in the DAS signal caused by noise, local anomalies, or minor fiber disturbances. Smoothing can employ Gaussian filtering or moving average filtering, convolving and averaging the amplitudes across the time and channel dimensions. This preserves the true energy spikes while suppressing random noise interference. After smoothing, the channel amplitude matrix becomes more continuous and stable along both the time and channel axes, facilitating the formation of intuitive distribution curves. After smoothing and normalizing the amplitude matrix, the system maps the matrix data to a three-dimensional coordinate system. The time axis is used as the X-axis, the channel sequence or spatial location as the Y-axis, and the amplitude intensity as the Z-axis. The magnitude of the amplitude is visually presented through three-dimensional surface mapping or color mapping. Regions with high amplitude form distinct peaks in the three-dimensional curve, while regions with low amplitude appear as flat areas or troughs, thus intuitively displaying the energy concentration locations of P-waves and S-waves and their propagation trends along the fiber array. For each candidate time period, the system connects consecutive amplitude peaks through three-dimensional interpolation or curve fitting to generate a smooth and continuous three-dimensional distribution curve, making the temporal concentration areas and spatial distribution trends of P-waves and S-waves more apparent.Ultimately, the generated three-dimensional distribution curves not only reflect the concentrated arrival locations of P-waves and S-waves in the time dimension, but also demonstrate their spatial distribution along the fiber array channel. By observing the spatiotemporal positions of the amplitude peaks in the curves, the arrival sequence and energy propagation paths of P-waves and S-waves can be accurately identified, thereby achieving high-precision positioning and energy distribution analysis of P-waves and S-waves in the DAS signal of the target area. Throughout the process, the channel amplitude matrix construction, normalization mapping, smoothing, and three-dimensional mapping were all strictly performed according to the monitoring time period and channel spatial distribution, ensuring that the final generated P-wave and S-wave arrival distribution curves are accurate, continuous, and usable for subsequent vibration mode analysis and phase identification.

[0109] like Figure 6 As shown, Figure 6 The present invention provides a flowchart of a lightweight DAS phase acquisition system based on sparse connectivity, including: a DAS signal segmentation and local time window generation module, used to segment and analyze DAS data in the target area, obtain the length of each segmentation window of the DAS data in the target area, segment the DAS data in the target area to form a DAS signal sequence in the target area, divide the DAS signal sequence in the target area to obtain each local time window, and dynamically adjust the morphological structural elements based on each local time window.

[0110] An enhanced signal feature sequence generation and dual-network input module is used to generate an enhanced signal feature sequence based on the DAS signal of the target region after the action of morphological structuring elements, and simultaneously input it into a parallel fusion network composed of convolution and attention mechanisms and a sparsely connected U-shaped network.

[0111] The parallel fusion network local and global vibration mode characterization module is used to accurately characterize and identify the vibration modes of the target region based on the parallel fusion network composed of convolution and attention mechanisms.

[0112] The sparsely connected U-shaped network feature map construction and channel activation monitoring module is used to construct and monitor the multi-channel feature maps corresponding to each convolution and downsampling layer in the convolutional branch of the sparsely connected U-shaped network. At the same time, it performs channel activation intensity analysis on the multi-channel feature maps corresponding to each convolution and downsampling layer and judges the execution of convolution and downsampling operations.

[0113] The multi-channel feature fusion and determination module is used to fuse the multi-channel feature maps output by the parallel fusion network composed of convolution and attention mechanisms and the sparse connection U-shaped network to determine the arrival candidate regions of P-waves and S-waves, and form the arrival distribution curves of P-waves and S-waves.

[0114] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0115] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device, equipment, and storage medium embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0116] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0117] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A lightweight DAS phase picking method based on sparse connectivity, characterized in that, The method includes: The target region DAS data is segmented and analyzed to obtain the length of each segmentation window. The target region DAS data is segmented to form the target region DAS signal sequence. The target region DAS signal sequence is divided to obtain each local time window. The morphological structural elements are dynamically adjusted based on each local time window. An enhanced signal feature sequence is generated based on the DAS signal of the target region after the action of morphological structural elements, and simultaneously input into a parallel fusion network composed of convolution and attention mechanisms and a sparse connection U-shaped network. The vibration patterns of the target region are accurately characterized and identified by a parallel fusion network composed of convolution and attention mechanisms. In the convolutional branch of the sparsely connected U-shaped network, multi-channel feature maps corresponding to each convolution and downsampling layer are constructed and monitored. At the same time, channel activation intensity analysis is performed on the multi-channel feature maps corresponding to each convolution and downsampling layer, and the execution of convolution and downsampling operations is judged. The multi-channel feature maps output by the parallel fusion network composed of convolution and attention mechanisms and the sparse connection U-shaped network are fused to determine the arrival candidate regions of P-waves and S-waves, forming the arrival distribution curves of P-waves and S-waves.

2. The lightweight DAS phase picking method based on sparse connectivity as described in claim 1, characterized in that, The specific process of segmenting the DAS data of the target area to form the DAS signal sequence of the target area is as follows: The target area DAS data is segmented and analyzed. The target area DAS data is initially segmented to obtain each first segmentation window of the target area DAS data. The temporal vibration response data along the optical fiber in each first segmentation window of the target area DAS data is obtained, including acoustic energy intensity, instantaneous rate of change of signal and signal waveform variance. The scale adjustment factor of each first segmentation window of the DAS data in the target region is obtained by fusing the sound wave energy intensity, the instantaneous rate of change of the signal, and the variance of the signal waveform. The scale adjustment factor of each first segmentation window of the DAS data in the target region is used to evaluate the activity level of the signal within each first segmentation window. The length duration of each segmentation window of the DAS data in the target region is obtained based on the scale adjustment factor of each first segmentation window of the DAS data in the target region. The first segmentation window of the DAS data in the target region is adjusted according to the length duration of each segmentation window of the DAS data in the target region, thereby segmenting the DAS data in the target region to form a DAS signal sequence in the target region.

3. The lightweight DAS phase picking method based on sparse connectivity as described in claim 1, characterized in that, The specific process of dynamically adjusting morphological structural elements based on each local time window is as follows: The target region DAS data is segmented to form a target region DAS signal sequence. It is divided into several local segments according to time to obtain each local time window. The temporal vibration response data along the optical fiber of the target region DAS data in each local time window is extracted to obtain the scale adjustment factor of each second local time window of the target region DAS data. The scale adjustment factor of each second local time window of the target region DAS data is used to evaluate the vibration mode stability and energy distribution change characteristics of the local signal at a fine-grained time scale. The morphological structural elements are dynamically adjusted according to the scale adjustment factor of each second local time window of the target region DAS data.

4. The lightweight DAS phase picking method based on sparse connectivity as described in claim 1, characterized in that, The specific process for accurately characterizing and identifying the vibration modes of the target area is as follows: The DAS enhanced signal of the target region obtained after morphological structuring is arranged in the original time order to form a continuous enhanced signal feature sequence. The enhanced signal feature sequence is input into the convolution branch, in which several groups of convolution kernels of different lengths are set, each convolution kernel corresponding to a specific time scale. Each convolution kernel is slid along the time axis on the sequence, and at each sliding position, the signal segment of the current position and its neighborhood is taken, multiplied element-wise with the convolution kernel weights and summed to obtain the local feature response value of each convolution kernel at the current position. The local feature response value of each convolution kernel at the current position is used to evaluate the degree of matching between the local features of the DAS signal of the target region and the convolution kernel weight pattern at that time point and its neighborhood. Local energy and waveform change data of the signal are obtained in the signal range covered by each convolution kernel when it slides to a certain time point, including the average value of the square of the DAS signal amplitude in the target region and the instantaneous rate of change of the signal. If the average value of the squared amplitude of the DAS signal in the target region and the instantaneous rate of change of the signal both exceed the corresponding range of the squared amplitude and the range of the instantaneous rate of change of the signal stored in the database, then the local feature response value of the convolution kernel at the current position corresponding to the signal range covered by the convolution kernel when it slides to a certain time point is corrected; otherwise, it is not necessary to correct the local feature response value of the convolution kernel at the current position.

5. The lightweight DAS phase picking method based on sparse connectivity as described in claim 1, characterized in that, The specific process of performing channel activation intensity analysis on the multi-channel feature maps corresponding to each convolutional and downsampling layer is as follows: A continuous sequence of enhanced signal features is input into the convolutional branch of a sparsely connected U-shaped network. Convolution and downsampling operations are performed in the encoding path of the U-shaped network. The spatiotemporal feature responses at each scale are extracted layer by layer to generate a multi-feature response matrix. The corresponding multi-channel feature map is automatically constructed in each downsampling layer and denoted as the multi-channel feature map corresponding to each downsampling layer. Extract the feature map automatically constructed for each downsampling layer, take the amplitude of the feature vector of each channel in the feature map automatically constructed for each downsampling layer point by point in the time dimension, and then sum the amplitudes and divide by the total number of feature vectors in the channel to obtain the average response value of each channel corresponding to each downsampling layer. The average response value of each channel corresponding to each downsampling layer is compared with the average response threshold of the channel corresponding to the downsampling layer stored in the database. If the average response value of a channel corresponding to a downsampling layer is lower than the average response threshold of the channel corresponding to the downsampling layer, the channel corresponding to the downsampling layer is marked as a low response channel and its characteristics are not retained in the skip connection. Otherwise, it is not necessary to mark the channel corresponding to the downsampling layer as a low response channel.

6. The lightweight DAS phase picking method based on sparse connectivity as described in claim 5, characterized in that, The specific process for determining whether to perform convolution and downsampling operations is as follows: Statistical analysis is performed on the multi-channel feature maps corresponding to each downsampling layer in both the time and channel dimensions to obtain global statistics for each downsampling layer, including channel mean, channel variance, time energy variance, and time-frequency energy entropy. The channel mean, channel variance, time energy variance, and time-frequency energy entropy are normalized and fused to obtain the comprehensive complexity index value of each downsampling layer. The comprehensive complexity index value of each downsampling layer is used to reflect the complexity and information density of the features of each downsampling layer. The overall complexity index value of each downsampling layer is compared with the overall complexity index threshold of the downsampling layer stored in the database. If the overall complexity index value of a downsampling layer is lower than the overall complexity index threshold of the downsampling layer, further convolution and downsampling operations are automatically stopped. If the overall complexity index value of a downsampling layer is higher than or equal to the overall complexity index threshold of the downsampling layer, further convolution and downsampling operations are continued.

7. The lightweight DAS phase picking method based on sparse connectivity as described in claim 1, characterized in that, The process of fusing the multi-channel feature maps output by the parallel fusion network composed of convolution and attention mechanisms and the sparsely connected U-shaped network is as follows: By arranging the multi-channel feature maps output by the parallel fusion of convolution and attention mechanisms according to the time and channel dimensions, a local vibration mode response sequence is obtained. At the same time, the multi-channel feature maps output by the sparsely connected U-shaped network are arranged according to the same time and channel dimensions to obtain a multi-scale global vibration mode sequence. The local vibration mode response sequence is matched point-by-point with the multi-scale global vibration mode sequence, and then linearly superimposed to obtain the fused multi-channel feature sequence.

8. The lightweight DAS phase picking method based on sparse connectivity as described in claim 4, characterized in that, The specific process for determining the candidate arrival regions of P-waves and S-waves is as follows: The fused multi-channel feature sequence is scanned point by point along the time axis to obtain the amplitude of each channel at each time point; A preset monitoring time period is set. The local feature response values ​​of each convolutional kernel at the current position are extracted in each monitoring time period. The local feature response values ​​of each convolutional kernel at the current position are plotted on the time axis to form a local feature response sequence. The average value of the amplitude of the local feature response sequence is calculated to obtain the global amplitude of each monitoring time period. The global amplitude of each monitoring time period is compared with the typical amplitude thresholds of P-waves and S-waves pre-stored in the database, and the frequency range of the local characteristic response sequence is determined. Then, it is jointly marked with the fiber optic channel to form the candidate P-wave and S-wave arrival areas containing time interval and channel information.

9. The lightweight DAS phase picking method based on sparse connectivity as described in claim 8, characterized in that, The specific process for forming the P-wave and S-wave arrival distribution curves is as follows: The time periods marked as candidate P-wave and S-wave arrival intervals and their corresponding amplitude information within each monitoring time period are arranged in chronological order. Combined with the spatial distribution of each optical fiber channel, a channel amplitude matrix is ​​constructed. The amplitude is normalized and mapped on the time axis according to the strength, and the areas with higher amplitude are highlighted to form a three-dimensional distribution curve of P-wave and S-wave arrival.

10. A system applying the lightweight DAS phase picking method based on sparse connectivity as described in any one of claims 1-9, characterized in that, include: The DAS signal segmentation and local time window generation module is used to segment and analyze the DAS data of the target region, obtain the length of each segmentation window of the DAS data of the target region, divide the DAS data of the target region into segments to form the DAS signal sequence of the target region, divide the DAS signal sequence of the target region to obtain each local time window, and dynamically adjust the morphological structural elements based on each local time window. An enhanced signal feature sequence generation and dual-network input module is used to generate an enhanced signal feature sequence based on the DAS signal of the target region after the action of morphological structuring elements, and simultaneously input it into a parallel fusion network composed of convolution and attention mechanisms and a sparse connection U-shaped network. The parallel fusion network local and global vibration mode characterization module is used to accurately characterize and identify the vibration modes of the target region based on the parallel fusion network composed of convolution and attention mechanisms. The sparse connection U-shaped network feature map construction and channel activation monitoring module is used to construct and monitor the multi-channel feature maps corresponding to each convolution and downsampling layer in the convolution branches of the sparse connection U-shaped network. At the same time, it performs channel activation intensity analysis on the multi-channel feature maps corresponding to each convolution and downsampling layer and judges the execution of convolution and downsampling operations. The multi-channel feature fusion and determination module is used to fuse the multi-channel feature maps output by the parallel fusion network composed of convolution and attention mechanisms and the sparse connection U-shaped network to determine the arrival candidate regions of P-waves and S-waves, and form the arrival distribution curves of P-waves and S-waves.

Images