Spatio-temporal dual-dimensional attention fusion method and system for das seismic phase picking

Abstract

The application discloses a space-time dual-dimensional attention fusion method and system for DAS seismic phase picking, and relates to the technical field of data processing. The space-time dual-dimensional attention fusion method and system for DAS seismic phase picking comprises the following steps: S1, acquiring DAS original vibration observation data, and performing channel direction and time direction reconstruction and amplitude normalization processing to screen a candidate seismic phase response area; S2, constructing a space propagation attention graph, generating a time first arrival attention graph, and identifying P-wave and S-wave arrival boundaries; S3, performing dual-dimensional gate mutual verification to generate a space-time fusion seismic phase feature; and S4, performing ridge line constraint correction to generate a channel-level seismic phase behavior label. The application effectively improves the picking accuracy of P-wave arrival time and S-wave arrival time in DAS data, and solves the problem that mixed modeling of the time dimension and the space dimension leads to the initial arrival being covered by strong energy response and the discontinuity of the cross-channel propagation ridge line.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a spatiotemporal two-dimensional attention fusion method and system for DAS phase picking. Background Technology

[0002] With the development of distributed optical fiber acoustic sensing (DAS) in microseismic monitoring, underground space monitoring, pipeline vibration monitoring, and seismic event sensing, the data acquired by DAS systems exhibits characteristics such as a large number of channels, high sampling frequency, significant differences in channel coupling, and complex environmental vibration interference. Current processing methods often use the DAS channel time matrix as a regular two-dimensional feature map input to a convolutional network or a unified self-attention network, simultaneously extracting temporal and spatial textures within the same feature space. However, they fail to separate and model the abrupt arrival process within the same channel from the continuous propagation process across channels.

[0003] For example, the invention patent with announcement number CN112884134B discloses a time-domain-based convolutional neural network model and its application for earthquake phase recognition, including: (1) dividing the earthquake waveform dataset into a training set, a validation set, and a test set; (2) combining a convolutional neural network (CNN) with a time-domain neural network (TCN) to construct an S-TCN model, inputting the training set into three consecutive convolutional blocks for feature learning, and then inputting the feature-learned data into a main TCN block and two consecutive sub-TCN blocks for total feature extraction of earthquake P-waves and S-waves and extraction of contextual information, and finally sending the outputs of the two consecutive sub-TCN blocks to two parallel branches of the GRU module and the TimeDistributed module for phase recognition and phase picking; (3) training the S-TCN model and validating it to obtain the phase recognition and picking results. This model retains more features of the earthquake signal data and has high accuracy in earthquake phase recognition.

[0004] In the existing technology, when dealing with near-field micro-vibrations, low signal-to-noise ratio P-waves, persistent tailwave response, and abnormal channel interference, existing systems tend to use subsequent high-amplitude regions or isolated channel noise as the basis for phase picking, resulting in P-wave arrival time lag, an increase in S-wave candidate peak values, and breaks in the arrival time curves between channels.

[0005] Therefore, in order to address the above problems, there is an urgent need for a spatiotemporal two-dimensional attention fusion method and system for DAS phase picking. Summary of the Invention

[0006] Technical problems to be solved

[0007] To address the shortcomings of existing technologies, this invention provides a spatiotemporal dual-dimensional attention fusion method and system for DAS phase picking, which solves the problems of first arrival being covered by strong energy response and discontinuity of cross-channel propagation ridges caused by mixed modeling of time and space dimensions.

[0008] Technical solution

[0009] To achieve the above objectives, the present invention provides the following technical solution: a spatiotemporal dual-dimensional attention fusion method for DAS phase picking, comprising: S1, acquiring raw DAS vibration observation data, reconstructing the channel direction and time direction and normalizing the amplitude, screening candidate phase response regions, and forming initial candidate response data; S2, constructing a spatial propagation attention map based on the initial candidate response data, generating a time first-arrival attention map, and identifying the arrival boundaries of P-waves and S-waves; S3, performing dual-dimensional gating mutual verification based on the spatial propagation attention map and the time first-arrival attention map, generating spatiotemporal fused phase features, and enhancing the spatiotemporal fused phase features; S4, outputting the arrival times of P-waves and S-waves based on the enhanced phase picking features, performing ridge constraint correction, generating channel-level phase behavior labels, and summarizing them to form a phase picking report.

[0010] Further, the specific process of acquiring raw DAS vibration observation data and performing channel direction and time direction reconstruction and amplitude normalization is as follows: Acquire raw DAS vibration observation data, which includes fiber channel number, channel physical location, channel spacing, sampling frequency, sampling point number, phase change, strain rate response value, original amplitude value, channel adjacency relationship, equipment sampling timestamp, event trigger timestamp, fiber connector location identifier, and channel valid status identifier; establish a spatial channel index based on the fiber channel number and channel physical location, and reconstruct the fiber acquisition results into a DAS spatiotemporal observation matrix according to the channel direction and time direction; construct DAS spatiotemporal reference data with a unified channel order and unified time coordinate based on the fiber channel number, channel physical location, sampling point number, and equipment sampling timestamp; generate a spatial adjacency matrix based on the channel physical location, channel spacing, and channel adjacency relationship; and perform statistical analysis on the background observation segments before the event trigger based on the DAS spatiotemporal reference data to extract channel background data. Background response, background energy fluctuations, adjacent channel response consistency, and anomalous peak distribution are encoded into channel state description vectors. These vectors are then input into a quality mapping unit to generate channel quality masks. Background mean de-meaning and amplitude normalization are applied to the DAS spatiotemporal reference data to eliminate channel static bias and unify the response scale of different channels. A short-time energy map is generated based on sliding window energy calculation. The normalized waveform is first-order differencing along the time direction to obtain the response change between adjacent sampling points. This is combined with gradient suppression within a local window to obtain a time gradient map. Normalized amplitude, short-time energy response, time gradient response, channel quality weights, and channel position encodings are concatenated into feature channels according to the same channel time coordinate. Dimensional unification and response compression are performed through a feature mapping unit to generate a standard DAS spatiotemporal observation tensor. The standard DAS spatiotemporal observation tensor, channel quality mask, short-time energy map, time gradient map, and spatial adjacency matrix are organized according to a unified channel time index to construct a spatiotemporal observation reference set.

[0011] Further, the specific process for screening candidate seismic phase response regions and forming initial candidate response data is as follows: Based on the spatiotemporal observation reference set, short-time energy maps, temporal gradient maps, standard DAS spatiotemporal observation tensors, and channel quality masks are read according to channel number and sampling time position; using the background reference window before the event trigger as a comparison benchmark, the short-time energy increment of the response observation window behind the candidate sampling point relative to the background reference window is calculated, and an energy rise judgment boundary is generated based on the energy dispersion of the background reference window; when the short-time energy increment meets the energy rise judgment boundary, the sampling point is recorded as an energy rise sampling point; based on the temporal gradient map, the gradient response, gradient direction consistency, and gradient duration within the continuous sampling range behind the candidate sampling point are calculated, and a gradient abrupt change judgment boundary is generated based on the gradient dispersion of the background reference window; when the gradient response behind the candidate sampling point meets the gradient abrupt change judgment boundary, and the gradient direction remains consistent within the continuous sampling range, the sampling point is recorded as a gradient abrupt change sampling point; for responses that only appear at a single sampling point and do not form a continuous gradient change, they are marked as isolated peaks. Peaks are eliminated; sampling points that simultaneously satisfy the energy rise criterion and gradient abrupt change criterion are marked as candidate time response points; adjacent channels are read according to the spatial adjacency matrix, and the allowable arrival time offset range of adjacent channels is determined according to the channel spacing and propagation speed range; if the arrival time difference between the candidate time response point of the target channel and the candidate time response point of the adjacent channel falls within the allowable arrival time offset range, and the two have consistent energy change, consistent gradient direction, and consistent channel quality within the window before and after the candidate arrival time, then they are determined to satisfy cross-channel continuity; if the candidate time response point lacks cross-channel continuity support from adjacent channels, it is marked as an isolated time peak; if the candidate time response point forms a continuous arrival relationship, a continuous energy relationship, and a same gradient direction relationship in adjacent channels, it is marked as a candidate seismic phase response point; according to the channel position and time position of the candidate seismic phase response point, a local spatiotemporal block containing the channel where the candidate point is located, adjacent channels, the front background reference window, the rear response observation window, and the corresponding channel quality mask is extracted from the standard DAS spatiotemporal observation tensor to construct the initial candidate response data.

[0012] Furthermore, the specific process of constructing the spatial propagation attention map based on the initial candidate response data is as follows: Based on the initial candidate response data, extract the response difference between adjacent channels, cross-channel arrival time offset, adjacent channel energy continuity, local propagation slope, and channel quality mask along the fiber channel direction; using the sampling time position of the candidate response point of the target channel as a reference, translate the sampling points of the local response segments of adjacent channels according to the cross-channel arrival time offset; construct the candidate propagation map from the candidate response points in the channel, and search along the fiber channel direction for candidate propagation paths that satisfy the conditions of continuous arrival time change, consistent energy continuity, and propagation slope change less than the slope change threshold, and identify the candidate propagation paths as candidate seismic ridges; backproject the spatial propagation attention weights of the candidate response points to the channel time coordinates corresponding to the DAS spatiotemporal reference data according to the channel number and sampling time position, and perform time delay compensation expansion in combination with the candidate seismic ridge and cross-channel arrival time offset to generate the spatial propagation attention map.

[0013] Further, the specific process of generating the first-arrival attention map and identifying the arrival boundaries of the P-wave and S-wave is as follows: Time-series analysis of waveform changes before and after the candidate response point is performed along the sampling time direction. Using the sampling time corresponding to the candidate response point as the center, a background reference window is extracted from the front, and a response observation window is extracted from the back. The background reference window and response observation window of the candidate response point are compared, and the features obtained from the comparison are input into the time response encoding unit to generate a first-arrival response description. Based on the first-arrival response description, the first-arrival attention weights of the candidate response points are adjusted. The first-arrival attention weights are jointly encoded with the features obtained from the comparison to form candidate time response features, which are then input into the P-wave first-arrival branch and the S-wave arrival branch. The candidate weights output from the two branches are backfilled into the channel time coordinates according to the channel number and sampling time position, generating the P-wave first-arrival attention map and the S-wave first-arrival attention map.

[0014] Furthermore, based on the spatial propagation attention map and the first-arrival attention map, the specific process of generating spatiotemporal fused phase features through two-dimensional gating mutual verification is as follows: Based on the P-wave and S-wave first-arrival attention maps, time attention response curves are generated along the time direction of each channel. Using the background reference window before triggering as a comparison benchmark, gradient judgment boundaries are generated according to the gradient dispersion. Sampling positions where the waveform gradient changes from the background fluctuation state to exceed the gradient judgment boundary and continues to change are found, and these sampling positions are used as candidate arrival points. When a candidate arrival point shows an abrupt boundary from the background segment to the phase segment in the time direction, the time candidate gate is opened, and the candidate arrival point is sent to the spatial closure gate. Spatial closure verification is performed on the candidate arrival point to determine whether it is located on a continuous propagation ridge. The candidate arrival points in the corresponding channel are backtracked and corrected. The first-arrival features, spatial propagation features, and original DAS response features corresponding to the candidate arrival points after mutual verification through the time candidate gate and spatial closure gate are weighted and fused to generate spatiotemporal fused phase features.

[0015] Further, the specific process for enhancing the spatiotemporal fusion phase features is as follows: The spatiotemporal fusion phase features are subjected to channel dimensionality reduction and local spatiotemporal block division to obtain dimensionality-reduced local response blocks; frequency domain transformation is performed on the dimensionality-reduced local response blocks to extract dominant frequency concentration, spectral entropy, narrowband stability, and transient broadband enhancement features; if the dimensionality-reduced local response block exhibits dominant frequency concentration and duration distribution that satisfies narrowband interference characteristics, it is marked as a narrowband interference candidate region, and its weight in fusion is reduced; if the dimensionality-reduced local response block exhibits broadband energy enhancement and is consistent with the spatial propagation direction, its weight in fusion is increased; the spatiotemporal fusion phase features after frequency domain noise suppression are input into a selective integration unit, and a combination of grouped feature remapping and correlation sparse selection is used to generate query features, key features, and value features, and candidate features effective for phase picking are screened based on feature correlation; the screened effective candidate features are mapped to the same channel time index of the standard DAS spatiotemporal observation tensor according to channel number, and weighted fusion is performed to generate enhanced phase picking features.

[0016] Furthermore, the specific process of outputting P-wave arrival times and S-wave arrival times based on the enhanced phase picking features and performing ridge constraint correction is as follows: Based on the enhanced phase picking features, output encoding is performed according to the channel time coordinates, and the encoding is sent to the P-wave output head and S-wave output head respectively for differential analysis to generate P-wave probability sequences and S-wave probability sequences; Based on the P-wave probability sequences and S-wave probability sequences, candidate P-wave arrival times and candidate S-wave arrival times in each channel are located respectively, and the confidence of the candidate arrival times is adjusted according to the relative arrival times of P-wave and S-wave; P-wave candidate arrival times that satisfy cross-channel continuity in adjacent channels are connected to form candidate P-wave ridges, and S-wave candidate arrival times that satisfy cross-channel continuity in adjacent channels are connected to form candidate S-wave ridges. Ridge constraint correction is performed on the candidate arrival times, and the P-wave arrival times, S-wave arrival times, P-wave confidence, and S-wave confidence of each channel are output.

[0017] Further, the specific process of generating channel-level phase behavior labels and summarizing them to form a phase picking report is as follows: If the P-wave confidence of a certain channel is lower than the confidence threshold, the channel is marked as a weak P-wave picking state; if the number of S-wave candidate peaks in a certain channel exceeds the range, or if the S-wave candidate points fall into the tail region of the coma, the channel is marked as an S-wave multi-peak interference state; if the channel quality mask of a certain segment shows an abnormality, and the corresponding phase lacks cross-channel continuity to the time point, the segment is marked as a channel abnormal influence state; if a certain segment simultaneously meets the consistency of the first arrival characteristics, spatial propagation characteristics, and original DAS response characteristics, it is marked as a valid phase propagation state; after the channel phase picking is completed, all channel-level phase behavior labels are summarized, and the phase picking results are verified by channel-level label convergence, propagation ridge consistency verification, and local waveform window verification to obtain a phase picking report.

[0018] Furthermore, the second aspect of the present invention provides a spatiotemporal two-dimensional attention fusion system for DAS phase picking, applied to a spatiotemporal two-dimensional attention fusion method for DAS phase picking, comprising: a DAS spatiotemporal observation acquisition and benchmarking module, used to acquire raw DAS vibration observation data, and perform channel direction and time direction reconstruction and amplitude normalization processing, screen candidate phase response regions, and form initial candidate response data; an attention separation and analysis module, used to construct a spatial propagation attention map based on the initial candidate response data, and generate a time first-arrival attention map, identifying the arrival boundaries of P-waves and S-waves; a two-dimensional gated mutual verification fusion module, used to perform two-dimensional gated mutual verification based on the spatial propagation attention map and the time first-arrival attention map, generate spatiotemporal fused phase features, and enhance the spatiotemporal fused phase features; and a phase ridge constraint output module, used to output the arrival times of P-waves and S-waves based on the enhanced phase picking features, perform ridge constraint correction, generate channel-level phase behavior labels, and summarize them to form a phase picking report.

[0019] Beneficial effects

[0020] The present invention has the following beneficial effects:

[0021] (1) In this invention, the fiber channel number, channel physical location, sampling point number, sampling time information, phase change, strain rate response value and original amplitude value are uniformly mapped to DAS spatiotemporal reference data through the DAS spatiotemporal observation acquisition and benchmarking module. The standard DAS spatiotemporal observation tensor is generated through background segment statistics, short-time energy encoding, time gradient extraction and channel quality masking, so that each channel time sampling point has amplitude response, energy change, time change, channel quality and spatial adjacency attributes at the same time. This reduces the impact of channel coupling difference, joint section anomaly and sampling time sequence difference on phase picking from the data input end.

[0022] (2) In this invention, channel quality masks are used to participate in spatial propagation attention calculation. Background noise level, background energy fluctuation, adjacent channel correlation degree, and abnormal peak ratio are encoded as channel quality weights, and the channel quality weights are used as modulation information for spatial propagation correlation calculation. For weakly coupled channels, connector interference channels, isolated high-response channels, and channels with insufficient adjacent channel consistency, the system reduces their contribution to propagation ridge judgment; for channels with stable background response and consistent changes with adjacent channels, the system enhances their spatial propagation support role, thereby reducing the misleading effect of abnormal channels on P-wave arrival time and S-wave arrival time.

[0023] (3) In this invention, by constructing candidate propagation maps and searching for the continuity of propagation ridges, candidate response points in each channel are used as candidate propagation nodes. Node connections are established based on time delay continuity, energy continuity, local propagation slope, and channel quality weights to form P-wave and S-wave candidate propagation maps. The system searches for candidate propagation paths with continuous connection weights along the fiber optic channel direction and corrects the candidate arrival times based on continuous length, adjacent arrival time difference smoothness, local slope changes, and cross-channel cumulative confidence. This ensures that the final output result no longer depends on the single-channel probability peak value but is constrained by the cross-channel propagation ridge, improving the spatial continuity of DAS phase picking results.

[0024] (4) In this invention, a two-dimensional gated mutual verification fusion structure is constructed by a time candidate gate and a spatial closure gate. After determining the candidate arrival point in the time first arrival attention map, the spatial propagation attention map is used to verify whether the candidate arrival point is located on a continuous propagation ridge. When the spatial propagation ridge exists but the local time boundary is unclear, the arrival position in the corresponding channel is backtracked and refined using the time first arrival attention map. This two-way mutual verification process enables the candidate phase point to simultaneously satisfy the first arrival boundary condition in the time direction and the propagation continuity condition in the spatial direction, reducing the probability that isolated time peaks, tail wave tail peaks, and spatial break points are output as phase arrival points.

[0025] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0026] Figure 1 This is a flowchart of the spatiotemporal two-dimensional attention fusion method for DAS phase picking in this invention;

[0027] Figure 2 This is a diagram of the spatiotemporal two-dimensional attention fusion system architecture for DAS phase picking in this invention;

[0028] Figure 3 This is a line graph showing the changes in the indicators of the detection channel of this invention;

[0029] Figure 4 This is a line graph showing the changes in the candidate detection point indicators of this invention.

[0030] Figure 5 This is a comparison diagram of the P-wave and S-wave initial arrival attention response in this invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figure 1 - Figure 5 This invention provides a technical solution: a spatiotemporal dual-dimensional attention fusion method for DAS phase picking, comprising: S1, acquiring raw DAS vibration observation data, reconstructing the channel direction and time direction and normalizing the amplitude, screening candidate phase response regions, and forming initial candidate response data; S2, constructing a spatial propagation attention map based on the initial candidate response data, generating a time first-arrival attention map, and identifying the arrival boundaries of P-waves and S-waves; S3, performing dual-dimensional gating mutual verification based on the spatial propagation attention map and the time first-arrival attention map, generating spatiotemporal fused phase features, and enhancing the spatiotemporal fused phase features; S4, outputting the arrival times of P-waves and S-waves based on the enhanced phase picking features, performing ridge constraint correction, generating channel-level phase behavior labels, and summarizing them to form a phase picking report.

[0033] Specifically, the process of acquiring raw DAS vibration observation data and performing channel and time direction reconstruction and amplitude normalization is as follows:

[0034] Acquire raw vibration observation data from DAS. The raw vibration observation data from DAS includes fiber channel number, channel physical location, channel spacing, sampling frequency, sampling point number, phase change, strain rate response value, raw amplitude value, channel adjacency relationship, equipment sampling timestamp, event trigger timestamp, fiber connector location identifier, and channel valid status identifier.

[0035] A spatial channel index is established based on the fiber channel number and channel physical location. The fiber acquisition results are reconstructed into a DAS spatiotemporal observation matrix according to the channel direction and time direction. Based on the fiber channel number, channel physical location, sampling point sequence number, and device sampling timestamp, DAS spatiotemporal reference data with a unified channel order and unified time coordinate is constructed. A spatial adjacency matrix is ​​generated according to the channel physical location, channel spacing, and channel adjacency relationship. Based on the DAS spatiotemporal reference data, statistical analysis is performed on the background observation segment before the event trigger. The background observation segment consists of continuous sampling points corresponding to a fixed background duration before the event trigger timestamp. The fixed background duration and the number of corresponding sampling points are obtained by converting the sampling frequency and are recorded together with the DAS spatiotemporal reference data. When there is a data gap before the event trigger, the segment with the smallest gradient variance among the continuous available sampling points before the event trigger is selected as the background observation segment. The start and end positions of the background observation segment are determined by the device sampling timestamp and the event trigger timestamp under a unified time coordinate.

[0036] Using the observation segment before the event trigger as the background segment, the background amplitude dispersion, background energy fluctuation, and abnormal peak ratio of each channel are calculated. The background amplitude dispersion is represented by the median absolute deviation of the amplitude sequence of the background segment; the background energy fluctuation is represented by the standard deviation of the short-time energy sequence within the sliding window in the background segment; and the abnormal peak ratio is represented by the ratio of the number of sampling points exceeding the statistical boundary of the background amplitude to the total number of sampling points in the background segment. The waveform correlation, energy change consistency, and phase change consistency between the target channel and adjacent channels in the same background segment are jointly encoded to extract the channel background response, background energy fluctuation, adjacent channel response consistency, and abnormal peak distribution, and encoded as a channel state description vector. The adjacent channel response consistency is jointly determined by the normalized cross-correlation coefficient, short-time energy sequence similarity, and phase difference consistency value between the target channel and the first-order adjacent channels in the spatial adjacency matrix in the same background segment. The channel state description vector is input into the quality mapping unit to generate a channel quality mask. The quality mapping unit then... A linear mapping layer and a sigmoid normalization layer are used. The input is a channel state description vector, and the output is a channel quality weight corresponding to a channel number. The channel quality weights form a channel quality mask. Background mean removal and amplitude normalization are performed on the DAS spatiotemporal reference data to eliminate channel static bias and unify the response scale of different channels. The amplitude normalization process uses the median value of the amplitude in the background segment of the channel as the bias term and the absolute deviation of the median amplitude in the background segment as the scale factor to map the original amplitude of each channel to the normalized amplitude. The abrupt boundary from the background segment to the phase segment is obtained by extracting the temporal gradient. A sliding time window is set on the normalized waveform of each channel. The window length is calculated by converting the sampling frequency and the preset energy statistics duration, and the sliding step size is calculated by converting the sampling frequency and the preset update duration. When the window boundary is insufficient, the endpoint sampling value is used to fill in the gap. The sum of squares of the amplitude, the envelope energy and the energy increment of the adjacent window are jointly calculated to obtain the energy response, and a short-time energy map is generated according to the channel time coordinate mapping.

[0037] Short-time energy maps are generated based on sliding window energy calculations. Each element in the short-time energy map is the normalized result of the normalized sum of squared amplitudes within the corresponding channel and sliding window. First-order difference is performed on the normalized waveform along the time direction to obtain the response change between adjacent sampling points. This is combined with gradient suppression of isolated spikes within a local window to obtain a time gradient map. The length of the local window is calculated from the sampling frequency and the preset gradient verification duration. When the gradient response of the target sampling point exceeds the background gradient statistical boundary, but its preceding and following adjacent sampling points do not form gradient changes in the same direction, the target sampling point is... Points are marked as isolated spikes, and the gradient response of the target sampling point is replaced with the median gradient value within the local window. Normalized amplitude, short-time energy response, temporal gradient response, channel quality weights, and channel position codes are concatenated into feature channels according to the same channel time coordinate, and dimension unification and response compression are performed through feature mapping units to generate a standard DAS spatiotemporal observation tensor. The standard DAS spatiotemporal observation tensor, channel quality mask, short-time energy map, temporal gradient map, and spatial adjacency matrix are organized according to a unified channel time index to construct a spatiotemporal observation reference set. The standard DAS spatiotemporal observation tensor uses channel number and sampling time position as two-dimensional indexes. The channel quality mask uses channel number as index and is associated with the corresponding channel in the standard DAS spatiotemporal observation tensor. The short-time energy map and temporal gradient map use channel number and sampling time position as indexes and correspond to the amplitude characteristics of the same sampling point, respectively. The spatial adjacency matrix uses the adjacency relationship between channel numbers as index and is used to identify the spatial connection relationship between channels. In the subsequent spatial propagation attention calculation and temporal first-arrival attention calculation, the standard DAS spatiotemporal observation tensor, channel quality mask, short-time energy map, temporal gradient map and spatial adjacency matrix are synchronously called according to the same channel number and sampling time position.

[0038] In this implementation scheme, by reconstructing the channel and time directions, statistically analyzing the background segments, generating channel quality masks, normalizing amplitudes, constructing short-time energy maps, and constructing time gradient maps from the raw DAS vibration observation data, the sampling data from different fiber optic channels can be expressed under a unified channel time index. This reduces the interference of channel static offset, coupling strength differences, background noise fluctuations, joint section anomalies, and isolated spikes on subsequent phase picking.

[0039] Specifically, the process of screening candidate seismic phase response regions to form initial candidate response data is as follows:

[0040] Based on the spatiotemporal observation benchmark set, short-time energy maps, temporal gradient maps, standard DAS spatiotemporal observation tensors, and channel quality masks are read according to channel number and sampling time position. A background reference window before the event trigger is used as the comparison benchmark. The background reference window and the background observation segment determined during the spatiotemporal observation benchmark construction phase use the same number of sampling points and the same statistical caliber. The end position of the background reference window is located before the event trigger timestamp, and the start position of the background reference window is determined by the inverse relationship between the end position and the number of sampling points. The short-time energy increment of the response observation window behind the candidate sampling point relative to the background reference window is calculated, and an energy rise judgment boundary is generated based on the energy dispersion of the background reference window. When the short-time energy increment meets the energy rise judgment boundary... The sampling point is recorded as the energy rise sampling point. The gradient response, gradient direction consistency, and gradient duration within the continuous sampling range behind the candidate sampling point are calculated based on the time gradient map. The continuous sampling range behind the candidate sampling point is obtained by converting the sampling frequency and the observation duration of the phase response. The observation duration of the phase response is obtained by statistically analyzing the continuous distribution from the first arrival of the P wave to the main energy enhancement segment in the training sample set. The gradient abrupt change judgment boundary is generated based on the gradient dispersion of the background reference window. When the gradient response behind the candidate sampling point meets the gradient abrupt change judgment boundary and the gradient direction remains consistent within the continuous sampling range, the sampling point is recorded as the gradient abrupt change sampling point. For responses that appear only at a single sampling point and do not form a continuous gradient change, they are marked as isolated spikes and removed.

[0041] Sampling points that simultaneously satisfy both energy rise and gradient change criteria are marked as candidate time response points. Adjacent channels are read based on the spatial adjacency matrix, where adjacent channels are determined by their physical location and spacing. When the fiber optic connector location marker or channel validity status marker near the target channel shows failure, or when an adjacent channel is marked as abnormal by a channel quality mask, fiber optic segments are divided based on the continuity of the fiber optic connector location marker, channel validity status marker, and physical distance. The valid channel that is in an adjacent fiber optic segment and has the closest physical distance to the target channel is read as the cross-segment channel. The allowable arrival time offset range for adjacent channels is determined based on the channel spacing and propagation speed range. If the candidate time response point of the target channel is close to the candidate time response point of an adjacent channel... If the arrival time difference falls within the allowable arrival time offset range, and the two have consistent energy changes, gradient directions, and channel quality within the window before and after the candidate arrival time, then they are determined to satisfy cross-channel continuity. If the candidate time response point lacks cross-channel continuity support from adjacent channels, it is marked as an isolated time peak. If the candidate time response point forms a continuous arrival relationship, an energy continuity relationship, and a gradient direction relationship in adjacent channels, it is marked as a candidate seismic phase response point. According to the channel location and time location of the candidate seismic phase response point, a local spatiotemporal block containing the channel where the candidate point is located, adjacent channels, cross-segment channels, the front background reference window, the rear response observation window, and the corresponding channel quality mask is extracted from the standard DAS spatiotemporal observation tensor to construct the initial candidate response data. The front background reference window and the background observation segment have the same length. The rear response observation window is calculated from the sampling frequency and the observation duration of the phase response. When the energy response within the rear response observation window falls back to the corresponding decision boundary of the background reference window, the fallback position is taken as the end position of the response observation window. If the end of the rear response observation window still satisfies the energy continuity relationship, the response observation window is extended to the maximum response observation length obtained from the training sample set. The local spatiotemporal block includes the channel where the candidate point is located, adjacent channels, the background segment before the candidate point, the subsequent response segment of the candidate point, and the corresponding channel quality mask.

[0042] like Figure 3 The graph shows the changes in the indicators of the detection channels. The green line represents the channel quality weight, which remains consistently above 0.7, dropping to its lowest point at CH105, then rising steadily after CH105 and returning to its peak at CH107. The blue background noise level and the orange background energy fluctuations show a similar trend, with both indicators gradually increasing from CH101 to CH105, reaching their peak at CH105, and then declining steadily from CH105 to CH107. The overall pattern shows that the higher the channel background noise and energy fluctuations, the lower the corresponding channel quality weight. CH105 is most susceptible to environmental interference and has the worst channel quality, while CH101 and CH107 have less environmental disturbance and the best acquisition quality. The acquisition quality of each channel can be graded and screened based on these three indicators.

[0043] In this implementation scheme, by extracting local spatiotemporal blocks, the candidate point's channel, adjacent channels, cross-segment channels, front background reference window, rear response observation window, and channel quality mask are synchronously input into subsequent modules. This enables spatial propagation attention and time arrival attention to call corresponding data within the same candidate response area, thereby improving the verifiability of the candidate seismic phase response area and the consistency of subsequent P-wave and S-wave boundary identification.

[0044] Specifically, the process of constructing the spatial propagation attention map based on the initial candidate response data is as follows:

[0045] Based on the initial candidate response data, the response difference between adjacent channels, cross-channel arrival time offset, adjacent channel energy continuity, local propagation slope, and channel quality mask are extracted along the fiber channel direction. The channel quality mask is derived from the quality mapping result of the aforementioned channel state description vector, with a value ranging from 0 to 1, and is used to represent the credibility of the corresponding channel's participation in spatial propagation judgment. Using the sampling time position of the candidate response point of the target channel as a reference, the sampling points of the local response segments of adjacent channels are shifted according to the cross-channel arrival time offset, so that the response segments belonging to the same phase in different channels are aligned. Then, the normalized correlation, amplitude change consistency, and gradient change consistency between the aligned target channel response segments and the response segments of adjacent channels are calculated, and the correlation contribution of abnormal channels is suppressed by combining the channel quality mask to obtain the spatial propagation correlation of the candidate response point. Among them, if adjacent channels still have similar waveforms, similar energy change directions, and continuous gradient change trends after time delay compensation, it indicates that the candidate response point has strong propagation correlation in the spatial direction. If adjacent channels only have isolated high-amplitude responses and the waveforms are inconsistent after time delay compensation, their spatial propagation correlation is reduced. Normalized correlation, amplitude variation consistency, and gradient variation consistency are all normalized and mapped to the 0 to 1 range. Spatial propagation correlation is calculated by combining the above three items with the quality weights of adjacent channels, with a value range of 0 to 1.

[0046] Based on the energy continuity of adjacent channels and the local propagation slope, candidate response points in the channel are used as candidate propagation nodes. Candidate propagation nodes with time delay continuity and energy continuity between adjacent channels are connected to form a candidate propagation map. In the candidate propagation map, propagation connection weights between nodes are generated based on whether the energy of adjacent channels continuously increases or decreases, whether the arrival time offset of adjacent channels changes smoothly, whether the local propagation slope remains consistent, and whether the quality weight of the connecting channels is reliable. Candidate propagation paths that meet the requirements of continuous arrival time change, consistent energy continuity, and propagation slope change less than the slope change threshold are searched along the fiber optic channel direction. These candidate propagation paths are identified as candidate phase ridges. If a candidate response point cannot form a continuous propagation path with the adjacent channels before and after it, or if its addition causes a local break, abrupt slope change, or interruption of energy continuity in the propagation path, the candidate response point is determined not to belong to a stable phase ridge, and its spatial propagation attention weight is reduced. The spatial propagation attention weight is obtained by fusing the spatial propagation correlation of the candidate response point, the propagation connectivity weight in the candidate propagation path, the corresponding channel quality weight, and the continuity score of the candidate seismic phase ridge. The value ranges from 0 to 1 and is used to indicate the degree to which the candidate response point belongs to the continuous seismic phase propagation path in the spatial direction.

[0047] If the candidate response points form a continuous arrival time in the spatial direction and the arrival time changes between adjacent channels remain smooth, then the spatial propagation attention weight of the candidate response points is increased; if the channel where the candidate response point is located is marked as an abnormal channel by the channel quality mask, or if the adjacent channels lack a corresponding propagation trend, then the spatial propagation attention weight of the candidate response points is decreased.

[0048] A blank attention mapping matrix with the same size as the DAS spatiotemporal reference data is established, and the channel number, sampling time position, and spatial propagation attention weight corresponding to each candidate response point are used as sparse mapping nodes. Based on the candidate phase ridge line to which the candidate response point belongs and the cross-channel arrival time offset, the spatial propagation attention weight of the candidate response point is back-projected to the channel time coordinates corresponding to the DAS spatiotemporal reference data according to the channel number and sampling time position. Each candidate propagation node in the candidate propagation map is mapped to the same channel number and the same sampling time position in the blank attention mapping matrix according to its bound channel number and sampling time position. When the sampling time position corresponding to the node after time delay compensation is a non-integer sampling point, linear interpolation is used. The spatial propagation attention weight of the node is assigned to two adjacent sampling time positions; and time delay compensation expansion is performed by combining candidate phase ridges and cross-channel arrival time offsets to ensure that adjacent candidate points on the same propagation ridge maintain continuous correspondence in the channel time coordinates; for cases where multiple candidate response points overlap in the same region, normalization fusion is performed based on the propagation correlation of candidate points, ridge connection weights, and channel quality masks to avoid isolated high-weight points directly covering continuous propagation areas; finally, the sparse attention mapping matrix after backprojection is locally smoothed and boundary preserved to ensure that the spatial propagation attention is continuously enhanced near the real phase ridge and rapidly attenuated at isolated noise points and spatial fault locations, generating a spatial propagation attention map.

[0049] like Figure 4 The chart shows the changes in the candidate detection point indicators. The blue line represents the cross-channel arrival time offset, with values ​​close to 0 and minimal fluctuations throughout, indicating stable and controllable arrival time differences between candidate points. The orange adjacent channel energy continuity and the green space propagation attention weight change trends are highly similar, with both indicators steadily increasing from R1 to R4, reaching a peak at R4, and gradually declining from R4 to R7, consistently remaining above 0.8. The data patterns indicate that the neighborhood energy correlation and spatial propagation characteristics at position R4 are optimal, effectively suppressing the cross-channel time offset error for all candidate points, and demonstrating excellent continuity in signal propagation characteristics between channels, providing a quantitative basis for candidate gate state discrimination and vibration source localization.

[0050] In this implementation scheme, by mapping candidate propagation nodes back to DAS spatiotemporal reference data according to channel number and sampling time position, a spatial propagation attention map consistent with the original channel time coordinates is formed, providing a spatial continuity basis for time arrival attention fusion, candidate arrival time point verification, and P-wave and S-wave ridge constraint output.

[0051] Specifically, the process of generating the initial arrival attention map and identifying the arrival boundaries of P and S waves is as follows:

[0052] Along the sampling time direction, a stable background segment, a candidate first arrival segment, a main energy enhancement segment, and a wake attenuation segment are extracted. The stable background segment represents the channel background state before the arrival of the seismic phase; the candidate first arrival segment represents the transition region from the background state to the seismic phase response state; the main energy enhancement segment represents the region where the seismic phase energy is significantly enhanced; and the wake attenuation segment represents the region of continuous attenuation after the main seismic phase response. A time-series analysis is performed on the waveform changes before and after the candidate response point along the sampling time direction: taking the sampling time corresponding to the candidate response point as the center, a background reference window is extracted before the candidate response point, and a response observation window is extracted after the candidate response point. The background reference window uses the same number of sampling points as the background observation segment in the spatiotemporal observation benchmark construction stage. With the same statistical caliber, the background reference window ends before the candidate response point; the response observation window begins at the sampling time corresponding to the candidate response point. The length of the response observation window is calculated from the sampling frequency and the observation duration of the phase response. When the short-time energy within the response observation window falls back to the energy judgment boundary corresponding to the background reference window, the fallback position is taken as the end position of the response observation window. The short-time energy, envelope amplitude, time gradient, and duration of the sustained response are compared between the background reference window and the response observation window at the candidate response point. The short-time energy, amplitude statistics, and waveform changes within the background reference window and the response observation window are calculated respectively. The increase in energy within the response observation window relative to the energy within the background reference window is taken as the energy rise. The degree of intensity is used to characterize whether a candidate response point corresponds to the energy initiation position of the seismic phase response from a stable background. The difference between the envelope amplitude within the response observation window and the corresponding amplitude feature within the background reference window is calculated to obtain the amplitude difference before and after the window, which characterizes whether the vibration intensity changes significantly before and after the candidate response point. The normalized waveform near the candidate response point is subjected to a first-order difference along the time direction, and combined with the gradient direction consistency of continuous sampling points within the local window, the waveform gradient change is obtained, which characterizes whether the candidate response point is the starting boundary of a continuous response, rather than an isolated spike. The energy retention time is statistically analyzed within the response observation window after the candidate response point to obtain continuous vibration characteristics, which are used to distinguish continuous vibrations after the arrival of the true seismic phase. Vibration response and instantaneous impulse noise: The energy rise, the difference in amplitude between the front and rear windows, the waveform gradient change, and the continuous vibration characteristics are concatenated according to the same candidate response point, and the compared features are input into the time response encoding unit to generate a time arrival response description. The time response encoding unit includes a one-dimensional time mapping layer, a normalization layer, and a sigmoid output layer. The input is the energy rise, the difference in amplitude between the front and rear windows, the waveform gradient change, and the continuous vibration characteristics corresponding to the same candidate response point. The output is a time arrival response description with a value range of 0 to 1. The time arrival response description is used to characterize whether the candidate response point has the arrival boundary characteristics of entering the P-wave or S-wave response segment from the background segment in the time direction.If a candidate response point exhibits an energy increase and abrupt gradient change relative to the preceding background segment, the initial arrival attention weight of the candidate response point is increased; if the candidate response point only exhibits a strong amplitude peak and lacks an initial boundary for transitioning from the background segment to the response segment, the initial arrival attention weight of the candidate response point is decreased. The initial arrival attention weight is obtained by normalizing the initial arrival response description, the degree of energy increase, waveform gradient changes, and continuous vibration characteristics, and its value ranges from 0 to 1.

[0053] The time-arrival attention weights of candidate response points are adjusted based on the time-arrival response description. These weights are then jointly encoded with the features obtained from comparison to form candidate time response features. These features are input into the P-wave arrival branch and the S-wave arrival branch. The P-wave arrival branch is used to enhance candidate response points near the background transition position in the event window. The S-wave arrival branch uses the time range after the P-wave candidate position as a constraint to enhance candidate response points that meet the energy enhancement and sustained response features. The candidate weights output from the P-wave arrival branch are backfilled into the channel time coordinates according to the channel number and sampling time position. Combined with local time smoothing and boundary preservation processing, a P-wave time-arrival attention map is generated. The candidate weights output from the S-wave arrival branch are backfilled into the channel time coordinates. Combined with the P-wave candidate position and local time continuity, constraint corrections are performed to generate an S-wave time-arrival attention map.

[0054] Table 1 shows the statistical table of channel vibration characteristic parameters. Quantitative analysis of energy, waveform, and time-domain weighting indicators was conducted for the seven detection channels: Channel CH102: energy increase 0.56, difference in amplitude between front and rear windows 0.34, waveform gradient change 0.41, continuous vibration characteristic 0.58, initial arrival attention weight 0.63, candidate gate status is open; CH103: energy increase 0.61, difference in amplitude between front and rear windows 0.38, waveform gradient change 0.48, continuous vibration characteristic 0.62, initial arrival attention weight 0.68, candidate gate open; CH104: energy increase 0.68, difference in amplitude between front and rear windows 0.45, waveform gradient change 0.56, continuous vibration characteristic 0.69, initial arrival attention weight 0.76, candidate gate open; CH105: energy increase... The following parameters were considered for CH106: Energy rise level 0.74, difference in amplitude between front and back windows 0.52, waveform gradient change 0.63, continuous vibration feature 0.73, initial arrival attention weight 0.83, candidate gate opened; CH107: Energy rise level 0.65, difference in amplitude between front and back windows 0.43, waveform gradient change 0.54, continuous vibration feature 0.64, initial arrival attention weight 0.72, candidate gate reviewed; CH108: Energy rise level 0.58, difference in amplitude between front and back windows 0.36, waveform gradient change 0.46, continuous vibration feature 0.59, initial arrival attention weight 0.66, candidate gate reviewed.

[0055] The data shows that the characteristic indicators of CH102 to CH105 gradually increase, and the vibration characteristics are significant, all of which meet the conditions for opening the candidate doors. The indicator of CH106 drops slightly, but still meets the opening criteria. The characteristic parameters of CH107 and CH108 continue to decrease, and the vibration signal characteristics are weak. They are judged to be in a verification state, and the graded start and stop control of the candidate doors of the channel can be completed based on the five quantitative indicators.

[0056] Table 1. Statistical Table of Channel Vibration Characteristic Parameters Detection

[0057] Channel number Energy rise level Difference in amplitude between front and rear windows Waveform gradient change Continuous vibration characteristics Attention weight at the beginning of time Candidate gate state CH102 0.56 0.34 0.41 0.58 0.63 Open CH103 0.61 0.38 0.48 0.62 0.68 Open CH104 0.68 0.45 0.56 0.69 0.76 Open CH105 0.74 0.52 0.63 0.73 0.83 Open CH106 0.71 0.49 0.6 0.7 0.79 Open CH107 0.65 0.43 0.54 0.64 0.72 Review CH108 0.58 0.36 0.46 0.59 0.66 Review

[0058] like Figure 5The figure shows a comparison of the attention responses to the first arrival of P-waves and S-waves. The horizontal axis represents time position, and the vertical axis represents the attention response value. The blue curve represents the attention response to the first arrival of P-waves, and the orange curve represents the attention response to the first arrival of S-waves. As can be seen from the figure, the first arrival attention response to P-waves forms a major response peak at an earlier time position, indicating that the P-wave first arrival branch can enhance the weak abrupt change boundary in the early part of the event window. The first arrival attention response to S-waves forms a major response peak after the P-wave response, indicating that the S-wave first arrival branch can further extract the energy-enhancing boundary after the candidate position of P-waves. The two curves show a separate distribution in time position, indicating that by using differentiated remapping of the P-wave and S-wave first arrival branches, it is possible to avoid mixing the weak first arrival boundary of P-waves with the subsequent strong response region of S-waves, thus providing a time-direction distinction for candidate arrival point determination and two-dimensional gating mutual verification.

[0059] In this implementation scheme, P-wave time first arrival attention map and S-wave time first arrival attention map are generated by P-wave first arrival branch and S-wave first arrival branch respectively, so that the early first arrival boundary of P-wave and the energy enhancement boundary of S-wave can be distinguished in time response, reducing the P-wave lag and S-wave boundary confusion caused by determining the arrival time of the seismic phase solely based on the peak value of strong amplitude.

[0060] Specifically, the process of generating spatiotemporally fused seismic phase features by performing two-dimensional gating mutual verification based on the spatial propagation attention map and the temporal first-arrival attention map is as follows:

[0061] Based on the P-wave and S-wave first-arrival attention maps, time attention response curves are generated along the time direction of each channel. Using the background reference window before triggering as a comparison benchmark, gradient judgment boundaries are generated according to the gradient dispersion. Local response regions are extracted from the time attention response curves. The peak position, the rising position of the peak front, and the corresponding time gradient response of the local response regions are jointly analyzed to determine candidate response intervals. Within the candidate response intervals, the waveform gradient is traced back from the attention peak position to find the sampling position where the waveform gradient changes from the background fluctuation state to exceed the gradient judgment boundary and maintains the same direction of change within the continuous sampling length. The sampling position is used as the candidate arrival time point. The continuous sampling length is calculated by converting the sampling frequency and the minimum duration of the phase response. The gradient judgment boundary adopts the same statistical caliber as the aforementioned gradient abrupt change judgment boundary.

[0062] When a candidate arrival point exhibits a sudden boundary in the temporal direction, transitioning from the background segment to the seismic phase segment, a temporal candidate gate is activated. The temporal candidate gate takes the initial arrival attention weight, energy rise rate, and gradient change as inputs, and generates a time-gated value through normalization mapping. The time-gated value ranges from 0 to 1, representing the confidence level that the candidate arrival point belongs to the initial arrival boundary in the temporal direction. The candidate arrival point is then sent to a spatial closure gate. Spatial closure verification is performed on the candidate arrival point to determine whether it is located on a continuous propagation ridge.

[0063] Spatial closure verification includes verification of the smoothness of arrival time changes in adjacent channels, verification of the continuity of local propagation slope, verification of cross-channel continuity, and verification of the consistency of channel quality mask. The spatial closure gate takes the spatial propagation attention map, candidate phase ridge, channel quality mask, and channel time position of candidate arrival time as input, determines whether the candidate arrival time falls within the channel time coordinate range corresponding to the candidate phase ridge, and generates a spatial closure value based on the continuity score of the candidate phase ridge. The spatial closure value ranges from 0 to 1 and is used to indicate the credibility of the candidate arrival time belonging to a continuous propagation path in the spatial direction.

[0064] If a candidate arrival point abruptly changes in the temporal direction but lacks propagation support in the spatial direction, the candidate arrival point is marked as an isolated time peak and the fusion weight is reduced. If a candidate arrival point forms a propagation ridge in the spatial direction but the temporal boundary is unclear, the time search range is narrowed in reverse according to the position of the spatial propagation ridge. The search proceeds in reverse along time to the earliest sampling position that satisfies the energy rise determination boundary and the gradient determination boundary, and this sampling position is used as the corrected candidate arrival point. The time search range is determined by the predicted arrival position of the candidate phase ridge and the allowable arrival offset range.

[0065] The candidate arrival times within the corresponding channel are retrospectively corrected. The time arrival characteristics, spatial propagation characteristics, and original DAS response characteristics corresponding to the candidate arrival times after mutual verification through time candidate gates and spatial closure gates are weighted and fused to generate spatiotemporal fused seismic phase characteristics. The weighting fusion weights are obtained by normalizing the product of the time gate value and the spatial closure value, so that the spatiotemporal fused seismic phase characteristics simultaneously carry information on abrupt changes in time arrival and continuous spatial propagation information.

[0066] In this implementation scheme, by weighted fusion of time arrival characteristics, spatial propagation characteristics and original DAS response characteristics, the generated spatiotemporal fused seismic phase characteristics simultaneously include single-channel arrival variation information, cross-channel propagation continuity information and original vibration response information, providing a consistent fusion characteristic basis for frequency domain noise suppression, selective integration and P-wave and S-wave arrival time output.

[0067] Specifically, the process of enhancing the spatiotemporal fusion seismic phase characteristics is as follows:

[0068] Channel dimensionality reduction and local spatiotemporal block partitioning are performed on the spatiotemporal fused seismic phase features to obtain dimensionality-reduced local response blocks. The channel dimensionality reduction target is the feature channel dimension of the spatiotemporal fused seismic phase features. A one-to-one convolution or linear mapping layer is used to map the original feature channel number to the dimensionality-reduced channel number. The dimensionality-reduced channel number is determined by the original feature channel number and the compression coefficient. The compression coefficient is jointly determined by the inference error and computational overhead on the training sample set and written into the model parameters. The local spatiotemporal blocks are divided according to the continuous channel range and the continuous time range. The channel range is determined by the number of adjacent channels in the spatial adjacency matrix, and the time range is calculated by the sampling frequency and the observation duration of the seismic phase response. Overlapping regions are set between adjacent local spatiotemporal blocks. When the boundary is insufficient, adjacent effective sampling points are used to fill in the gaps.

[0069] A frequency domain transformation is performed on the reduced-dimensional local response block to extract the dominant frequency concentration, spectral entropy, narrowband stability, and transient broadband enhancement features. The frequency domain transformation uses short-time Fourier transform, and the time window length of the short-time Fourier transform is consistent with the time length of the local spatiotemporal block. The frequency resolution is determined by the sampling frequency and the number of sampling points in the time window. The dominant frequency concentration is the proportion of the maximum frequency band energy to the total frequency band energy. The spectral entropy is the information entropy of the normalized frequency band energy distribution. The narrowband stability is the proportion of the dominant frequency falling into the same frequency band within a continuous time window. The transient broadband enhancement feature is the normalized increment of the broadband energy of the response observation window relative to the broadband energy of the background reference window.

[0070] If the reduced local response block exhibits a concentration of dominant frequencies and a duration distribution that satisfies the narrowband interference characteristics, a narrowband interference score is generated based on the dominant frequency concentration, the inverse value of the spectral entropy, and the narrowband stability. A frequency domain suppression coefficient is then generated based on the narrowband interference score, and this frequency domain suppression coefficient is applied to the fusion feature weights corresponding to the reduced local response block. The dominant frequency concentration determination boundary, the spectral entropy determination boundary, and the narrowband stability determination boundary are obtained statistically from the background observation segments and the narrowband interference samples in the training sample set.

[0071] If the reduced local response block exhibits broadband energy enhancement and is consistent with the spatial propagation direction, its weight in the fusion process is increased. Broadband energy enhancement is determined by transient broadband enhancement features, and spatial propagation direction consistency is determined by the consistency of the local propagation slope of the spatial propagation attention map and the candidate seismic phase ridge. A broadband enhancement score is generated based on the transient broadband enhancement features and spatial propagation direction consistency, and a frequency domain enhancement coefficient is generated based on the broadband enhancement score. The frequency domain suppression coefficient and the frequency domain enhancement coefficient are then updated with fusion weights under the same normalized weighting framework.

[0072] The spatiotemporal fused seismic phase features, after frequency domain noise suppression, are input into a selective integration unit. A combination of grouped feature remapping and correlation sparse selection is used to generate query features, key features, and value features. Grouped feature remapping groups the spatiotemporal fused seismic phase features according to channel segments, time segments, and candidate seismic phase ridgeline affiliations. Query features are obtained by mapping the initial arrival time attention weight, spatial propagation attention weight, and channel quality weight of candidate response points, representing the retrieval needs of the current candidate response point for seismic phase discrimination features. Key features are obtained by mapping the frequency domain suppression coefficient, frequency domain enhancement coefficient, candidate seismic phase ridgeline continuity score, and channel quality weight of candidate feature segments, representing the degree of matching between candidate feature segments and effective seismic phase propagation. Value features are obtained by mapping the original fused features and the frequency-modulated fused features of the corresponding candidate feature segments, preserving the response content of the candidate feature segments. Based on feature correlation, effective candidate features for seismic phases are selected. The normalized correlation between query features and key features is used to calculate the candidate feature correlation score. The candidate feature correlation score is then combined with channel quality weight, candidate seismic phase ridge continuity score and frequency domain suppression coefficient to obtain a sparse selection score. When the sparse selection score meets the screening boundary determined by the distribution of effective seismic phase candidate feature scores in the training sample set, the corresponding candidate feature is retained. When the sparse selection score does not meet the screening boundary, the corresponding candidate feature is suppressed.

[0073] The selected valid candidate features are mapped to the same channel time index of the standard DAS spatiotemporal observation tensor according to the channel number, and then weighted and fused to generate enhanced phase picking features. The weighting fusion weights are jointly determined by sparse selection score, spatial propagation attention weight, first arrival attention weight, and frequency domain modulation weight, so that the enhanced phase picking features simultaneously retain the original DAS response information, propagation ridge information, first arrival boundary information, and frequency domain noise reduction information.

[0074] In this implementation scheme, the selected effective candidate features are mapped to the same channel time index of the standard DAS spatiotemporal observation tensor and then weighted and fused. This allows the enhanced phase picking features to retain the original DAS response information and superimpose the phase discrimination information after frequency domain noise suppression and correlation screening, providing a stable enhanced feature basis for the P-wave arrival time, S-wave arrival time and candidate phase ridge output.

[0075] Specifically, the process of outputting the P-wave arrival time and S-wave arrival time based on the enhanced phase picking characteristics and performing ridge constraint correction is as follows:

[0076] Based on the enhanced phase picking features, output encoding is performed according to the channel time coordinates. The output encoding includes local temporal convolution, spatial adjacency aggregation, and point-based probability mapping. Local temporal convolution is used to extract the temporal boundary features before and after the sampling point in each channel. Spatial adjacency aggregation is used to aggregate the propagation continuity features of adjacent channels based on the spatial adjacency matrix. Point-based probability mapping is used to map each channel sampling point to a phase candidate score. The feature size of the output encoding is consistent with the channel time coordinates of the DAS spatiotemporal reference data, and each channel sampling point corresponds to an output encoding feature.

[0077] By employing local temporal convolution, channel orientation feature aggregation, and point-based classification mapping, the temporal boundary information within each channel and the propagation continuity information between adjacent channels are preserved. These are then fed into P-wave and S-wave output heads for differential analysis. The P-wave output head uses the first set of mapping parameters, focusing on temporal gradient changes, energy initiation changes, and P-wave first-arrival attention weights to generate P-wave scores. The S-wave output head uses the second set of mapping parameters, constrained by the time range after the P-wave candidate position, and combines principal energy enhancement features, sustained response features, and S-wave first-arrival attention weights to generate S-wave scores. A probability normalization layer then converts the scores of each sampling point into probability values, which are arranged according to channel number and sampling time order to generate P-wave and S-wave probability sequences. The P-wave probability sequence represents the probability that each sampling point within each channel belongs to the P-wave arrival time, and the S-wave probability sequence represents the probability that each sampling point within each channel belongs to the S-wave arrival time.

[0078] Based on the P-wave and S-wave probability sequences, candidate arrival times for P-waves and S-waves are located within each channel. Candidate arrival times are determined by the local extrema in the probability sequences and the corresponding temporal gradient boundaries, avoiding the use of only the maximum probability value as the arrival time. The confidence level of candidate arrival times is adjusted according to the relative arrival times of P-waves and S-waves. The relative arrival times of P-waves and S-waves are as follows: within the same channel, the candidate arrival time of S-waves is located after the candidate arrival time of P-waves, and the time difference between them lies within the allowable relative arrival time range determined by the statistical distribution of P-wave and S-wave arrival times in the training sample set. The confidence level of candidate arrival times of S-waves that are earlier than the candidate arrival time of P-waves or do not conform to the relative arrival times of P-waves and S-waves is reduced. The confidence level adjustment is obtained by normalizing and multiplying the original probability of the candidate arrival time, the relative arrival time consistency score, and the candidate ridge continuity score. The adjusted confidence level ranges from 0 to 1.

[0079] Candidate P-wave arrival points that satisfy cross-channel continuity in adjacent channels are connected to form candidate P-wave ridges. Candidate S-wave arrival points that satisfy cross-channel continuity in adjacent channels are connected to form candidate S-wave ridges. Ridge constraint corrections are applied to the candidate arrival points. Isolated candidate arrival points that are not connected to candidate ridges and lack propagation support from adjacent channels are removed from the candidate arrival point set. For candidate arrival points that, after being added to candidate ridges, cause discontinuities in arrival time differences between adjacent channels or local propagation slope changes exceeding the slope change threshold, they are selected based on their proximity to the ridge in the candidate probability sequence within the same channel. The candidate points for the predicted arrival time position are replaced; for channel segments in the candidate ridge line where the local arrival time fluctuation exceeds the boundary of the continuity judgment of the arrival time difference of the adjacent channel, the candidate arrival time points of the channel segment are smoothed and corrected according to the arrival time change trend of the adjacent channels before and after the candidate ridge line; for candidate ridge lines that still meet the cross-channel continuity, local slope change constraints and frequency domain feature consistency after correction, their cross-channel cumulative confidence is updated, and the P-wave arrival time, S-wave arrival time, P-wave confidence and S-wave confidence of each effective channel are output based on the corrected candidate P-wave ridge line and candidate S-wave ridge line.

[0080] In this implementation scheme, candidate arrival times of adjacent channels are constrained and corrected by candidate P-wave ridges and candidate S-wave ridges. Isolated candidate points lacking propagation support are deleted, and candidate points that disrupt the continuity of local propagation slope are replaced. Channel segments where local arrival time fluctuations exceed the continuity determination boundary are smoothed and corrected, so that the output P-wave arrival time, S-wave arrival time, P-wave confidence, and S-wave confidence simultaneously satisfy the single-channel time boundary determination and the cross-channel propagation continuity determination.

[0081] Specifically, the process of generating channel-level seismic phase behavior labels and summarizing them to form a seismic phase picking report is as follows:

[0082] The confidence threshold, the range of candidate peak counts, and the channel anomaly detection criteria are obtained statistically from the P-wave arrival time deviation, S-wave candidate peak distribution, channel quality mask distribution, and ridge continuity score in the training sample set or historical labeled sample set. The range of candidate peak counts is the upper limit of the number of S-wave candidate peaks allowed to be retained within the same channel and the same S-wave candidate search window. The S-wave candidate search window is determined by the allowable relative arrival time range after the P-wave candidate arrival time. The upper limit of the number of candidate peaks is obtained statistically from the distribution of candidate peak counts near the effective S-wave arrival time in the training sample set.

[0083] If the confidence level of the P-wave in a certain channel is lower than the confidence level threshold, the channel is marked as a weak P-wave pickup state. If the number of S-wave candidate peaks in a certain channel exceeds the range, or if the S-wave candidate points fall into the tail region, the channel is marked as a multi-peak S-wave interference state. The tail region is the time interval after the main candidate peak of the S-wave, where the short-term energy continues to decrease and the time gradient no longer meets the initial arrival boundary determination condition. The starting position of the tail region is determined by the position of the main candidate peak of the S-wave, and the ending position is determined by the position where the short-term energy falls back to the background reference window energy determination boundary.

[0084] If the channel quality mask of a certain segment shows an anomaly and the corresponding seismic phase lacks cross-channel continuity to the time point, the segment is marked as being in an abnormal channel influence state. The channel quality mask ranges from 0 to 1. When the channel quality mask is lower than the channel quality judgment boundary, the corresponding channel is marked as being in an abnormal state. The channel quality judgment boundary is obtained by statistically analyzing the channel quality weight distribution of normal and abnormal channels in the training sample set or the historical labeled sample set.

[0085] If a segment simultaneously satisfies the consistency of its first-arrival characteristics, spatial propagation characteristics, and original DAS response characteristics, it is marked as a valid phase propagation state. Consistency of first-arrival characteristics means that candidate arrival points within the segment meet the energy rise judgment boundary, gradient abrupt change judgment boundary, and first-arrival attention weight judgment boundary. Consistency of spatial propagation characteristics means that candidate arrival points within the segment meet the cross-channel continuity relationship, local propagation slope change constraint, and candidate ridge continuity score judgment boundary. Consistency of frequency domain characteristics means that candidate response blocks within the segment do not meet the narrowband interference judgment condition and meet the transient broadband enhancement judgment condition. When all three consistency conditions are met simultaneously, the segment is marked as a valid phase propagation state. After channel phase picking is completed, all channel-level phase behavior labels are summarized. Channel-level label aggregation, propagation ridge consistency verification, and local waveform window verification are used to verify the phase picking results, resulting in a phase picking report.

[0086] In this implementation plan, by defining the tail region, channel anomaly status, and effective phase propagation status, the impact of persistent tail response, low-quality channels, and isolated candidate peaks on the report is reduced. Through channel-level label aggregation, propagation ridge consistency verification, and local waveform window verification, the single-channel picking results are jointly verified with the cross-channel propagation relationship, so that the phase picking report can simultaneously reflect the time result, confidence result, channel quality status, ridge continuity status, and anomalous impact section.

[0087] Specifically, the second aspect of the present invention provides a spatiotemporal two-dimensional attention fusion system for DAS phase picking, applied to a spatiotemporal two-dimensional attention fusion method for DAS phase picking, comprising: a DAS spatiotemporal observation acquisition and benchmarking module, used to acquire raw DAS vibration observation data and perform channel direction and time direction reconstruction and amplitude normalization processing. The channel direction and time direction reconstruction includes establishing a spatial channel index and unifying the time sampling axis. The amplitude normalization processing includes eliminating channel static bias and unifying the channel response scale based on the background segment amplitude statistics, screening candidate phase response areas, and forming initial candidate response data; an attention separation and analysis module, used to construct a spatial propagation attention map based on the initial candidate response data and generate a time arrival attention map, identifying the arrival boundaries of P-waves and S-waves; the time arrival attention map includes a P-wave time arrival attention map and an S-wave time arrival attention map, the P-wave time arrival attention map being generated by the P-wave time arrival branch, and the S-wave time arrival attention map being generated by the S-wave time arrival branch. The dual-gated mutual verification fusion module is used to perform dual-gated mutual verification based on spatial propagation attention maps and time first-arrival attention maps to generate spatiotemporal fused seismic phase features and enhance these features. Dual-gated mutual verification includes generating time-gated values ​​based on the time first-arrival attention map and generating spatial closure values ​​based on the spatial propagation attention map and candidate seismic phase ridges. The time-gated values ​​and spatial closure values ​​are normalized and fused to obtain dual-dimensional fusion weights, and then weighted and fused based on these weights for the time first-arrival features, spatial propagation features, and original DAS response features. Enhancement of the spatiotemporal fused seismic phase features includes frequency domain noise suppression and selective feature integration. The phase ridge constraint output module is used to output P-wave arrival times and S-wave arrival times based on the enhanced phase picking characteristics, perform ridge constraint correction, generate channel-level phase behavior labels, and summarize them into a phase picking report. The P-wave arrival times and S-wave arrival times are generated by the P-wave output head and the S-wave output head, respectively. The ridge constraint correction includes connecting candidate arrival times that satisfy the propagation continuity relationship in adjacent channels to form candidate P-wave ridges and candidate S-wave ridges. Based on the continuous length of the candidate ridges, the continuity of adjacent arrival time differences, local slope changes, and cross-channel cumulative confidence, isolated candidate arrival times, candidate arrival times that deviate from the ridge direction, and candidate arrival times that disrupt the ridge continuity are deleted, replaced, or smoothed. The channel-level phase behavior labels include P-wave weak picking status, S-wave multi-peak interference status, channel anomaly influence status, and effective phase propagation status.

[0088] In this implementation scheme, candidate arrival times are mutually verified and fused using time gating values ​​and spatial closure values, so that the spatiotemporal fused seismic phase characteristics are simultaneously constrained by time arrival characteristics and spatial propagation characteristics. Through frequency domain noise suppression processing, selective feature integration, and seismic ridge constraint correction, P-wave arrival time, S-wave arrival time, confidence results, and channel-level seismic phase behavior labels can be generated in the same system link, forming a seismic phase picking report that includes picking results, ridge continuity status, and anomalous influence status.

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0090] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A spatiotemporal two-dimensional attention fusion method for DAS phase picking, characterized in that, Includes the following steps: S1. Obtain the raw vibration observation data from DAS, and perform channel direction and time direction reconstruction and amplitude normalization processing to screen candidate seismic phase response zones and form initial candidate response data. S2, based on the initial candidate response data, construct a spatial propagation attention map and generate a temporal arrival attention map to identify the arrival boundaries of P-waves and S-waves; S3, based on the spatial propagation attention map and the time first arrival attention map, performs two-dimensional gating mutual verification to generate spatiotemporal fusion seismic phase features and enhances the spatiotemporal fusion seismic phase features; S4 outputs the arrival times of P-waves and S-waves based on the enhanced phase picking features, performs ridge constraint correction, generates channel-level phase behavior labels, and summarizes them to form a phase picking report.

2. The spatiotemporal two-dimensional attention fusion method for DAS phase picking according to claim 1, characterized in that: The specific process of acquiring raw DAS vibration observation data and performing channel direction and time direction reconstruction and amplitude normalization is as follows: Acquire raw vibration observation data from DAS. The raw vibration observation data from DAS includes fiber channel number, channel physical location, channel spacing, sampling frequency, sampling point number, phase change, strain rate response value, raw amplitude value, channel adjacency relationship, equipment sampling timestamp, event trigger timestamp, fiber connector location identifier, and channel valid status identifier. A spatial channel index is established based on the fiber channel number and channel physical location, and the fiber acquisition results are reconstructed into a DAS spatiotemporal observation matrix according to the channel direction and time direction. Based on fiber channel number, channel physical location, sampling point sequence number and device sampling timestamp, DAS spatiotemporal reference data with unified channel order and unified time coordinate is constructed. A spatial adjacency matrix is ​​generated based on the physical location of the channels, the channel spacing, and the channel adjacency relationships; Based on DAS spatiotemporal reference data, statistical analysis is performed on background observation segments before the event is triggered to extract channel background response, background energy fluctuation, adjacent channel response consistency and abnormal peak distribution, and encode them into channel state description vectors; the channel state description vectors are input into the quality mapping unit to generate channel quality masks; background mean removal and amplitude normalization are performed on the DAS spatiotemporal reference data to eliminate channel static bias and unify the response scale of different channels; Short-time energy maps are generated based on sliding window energy calculations. First-order differences are performed on the normalized waveform along the time direction to obtain the response changes between adjacent sampling points. This is combined with gradient suppression within the local window to obtain a time gradient map. Normalized amplitude, short-time energy response, time gradient response, channel quality weights, and channel position codes are then concatenated into feature channels according to the same channel time coordinate. Dimensional unification and response compression are performed using feature mapping units to generate a standard DAS spatiotemporal observation tensor. The standard DAS spatiotemporal observation tensor, channel quality mask, short-time energy map, time gradient map, and spatial adjacency matrix are organized according to a unified channel time index to construct a spatiotemporal observation reference set.

3. The spatiotemporal two-dimensional attention fusion method for DAS phase picking according to claim 1, characterized in that: The specific process of screening candidate seismic phase response regions and forming initial candidate response data is as follows: Based on the spatiotemporal observation reference set, short-time energy map, temporal gradient map, standard DAS spatiotemporal observation tensor and channel quality mask are read according to channel number and sampling time position; Using the background reference window before the event is triggered as a comparison benchmark, the short-term energy increment of the response observation window behind the candidate sampling point relative to the background reference window is calculated, and an energy rise judgment boundary is generated based on the energy dispersion of the background reference window. When the short-term energy increment meets the energy rise determination boundary, the sampling point is recorded as the energy rise sampling point; The gradient response, gradient direction consistency, and gradient duration within the continuous sampling range following the candidate sampling point are calculated based on the time gradient plot. Gradient abrupt change judgment boundaries are generated based on the gradient dispersion of the background reference window. When the gradient response following the candidate sampling point satisfies the gradient abrupt change judgment boundaries and the gradient direction remains consistent within the continuous sampling range, the sampling point is recorded as a gradient abrupt change sampling point. For responses that appear only at a single sampling point and do not form a continuous gradient change, they are marked as isolated spikes and removed. Sampling points that simultaneously satisfy both the energy rise criterion and the gradient change criterion are marked as candidate time response points; Adjacent channels are read based on the spatial adjacency matrix, and the allowable arrival time offset range of adjacent channels is determined based on the channel spacing and propagation speed range. If the arrival time difference between the candidate time response point of the target channel and the candidate time response point of the adjacent channel falls within the allowable arrival time offset range, and the two have consistent energy change, consistent gradient direction, and consistent channel quality within the window before and after the candidate arrival time, then they are determined to satisfy cross-channel continuity. If the candidate time response point lacks cross-channel continuity support from adjacent channels, it is marked as an isolated time peak. If the candidate time response point forms a continuous arrival relationship, a continuous energy relationship, and a common gradient direction relationship in adjacent channels, it is marked as a candidate seismic phase response point. According to the channel position and time position of the candidate seismic phase response point, a local spatiotemporal block containing the channel where the candidate point is located, adjacent channels, the front background reference window, the rear response observation window, and the corresponding channel quality mask is extracted from the standard DAS spatiotemporal observation tensor to construct the initial candidate response data.

4. The spatiotemporal two-dimensional attention fusion method for DAS phase picking according to claim 1, characterized in that: The specific process of constructing the spatial propagation attention map based on the initial candidate response data is as follows: Based on the initial candidate response data, the response difference between adjacent channels, cross-channel arrival time offset, adjacent channel energy continuity, local propagation slope and channel quality mask are extracted along the fiber channel direction. Based on the sampling time position of the candidate response points of the target channel, the sampling points of the local response segments of adjacent channels are shifted according to the cross-channel arrival time offset; the candidate response points in the channel are constructed into a candidate propagation map, and the candidate propagation path in the candidate propagation map that meets the requirements of continuous arrival time change, consistent energy continuity and propagation slope change less than the slope change threshold is searched along the fiber channel direction. The candidate propagation path is identified as the candidate phase ridge line. The spatial propagation attention weights of candidate response points are backprojected to the corresponding channel time coordinates of the DAS spatiotemporal reference data according to the channel number and sampling time position, and time delay compensation extension is performed by combining candidate seismic ridge lines and cross-channel arrival time offsets to generate a spatial propagation attention map.

5. The spatiotemporal two-dimensional attention fusion method for DAS phase picking according to claim 1, characterized in that: The specific process of generating an initial arrival attention map and identifying the arrival boundaries of P-waves and S-waves is as follows: Timing analysis is performed on the waveform changes before and after the candidate response point along the sampling time direction: Taking the sampling time corresponding to the candidate response point as the center, a background reference window is extracted in front and a response observation window is extracted behind. The background reference window and the response observation window of the candidate response point are compared, and the features obtained from the comparison are input into the time response encoding unit to generate the time arrival response description. The time arrival attention weights of the candidate response points are adjusted based on the time arrival response descriptions. The time arrival attention weights are jointly encoded with the features obtained from the comparison to form candidate time response features, which are then input into the P-wave arrival branch and the S-wave arrival branch. The candidate weights output from the two branches are backfilled into the channel time coordinates according to the channel number and sampling time position to generate the P-wave time arrival attention map and the S-wave time arrival attention map.

6. The spatiotemporal two-dimensional attention fusion method for DAS phase picking according to claim 1, characterized in that: The specific process of generating spatiotemporally fused seismic phase features by performing two-dimensional gating mutual verification based on spatial propagation attention maps and time first-arrival attention maps is as follows: Based on the P-wave first-arrival attention map and the S-wave first-arrival attention map, time attention response curves are generated along the time direction of each channel. Using the background reference window before triggering as a comparison benchmark, gradient judgment boundaries are generated according to the gradient dispersion. Sampling positions where the waveform gradient changes from the background fluctuation state to exceed the gradient judgment boundary and continues to change are found, and these sampling positions are used as candidate arrival points. When a candidate arrival point shows an abrupt boundary from the background segment to the phase segment in the time direction, the time candidate gate is opened, and the candidate arrival point is sent to the spatial closure gate. Spatial closure verification is performed on the candidate arrival point to determine whether it is located on a continuous propagation ridge. The candidate arrival points in the corresponding channel are backtracked and corrected. The first-arrival features, spatial propagation features, and original DAS response features corresponding to the candidate arrival points after mutual verification by the time candidate gate and the spatial closure gate are weighted and fused to generate spatiotemporally fused phase features.

7. The spatiotemporal two-dimensional attention fusion method for DAS phase picking according to claim 1, characterized in that: The specific process for enhancing the spatiotemporal fusion seismic phase characteristics is as follows: The spatiotemporal fusion seismic phase features are subjected to channel dimensionality reduction and local spatiotemporal block partitioning to obtain dimensionality-reduced local response blocks. Frequency domain transformation is performed on these dimensionality-reduced local response blocks to extract dominant frequency concentration, spectral entropy, narrowband stability, and transient broadband enhancement features. If the dimensionality-reduced local response block exhibits dominant frequency concentration and duration distribution that satisfies narrowband interference characteristics, it is marked as a narrowband interference candidate region, and its weight in fusion is reduced. If the dimensionality-reduced local response block exhibits broadband energy enhancement consistent with the spatial propagation direction, its weight in fusion is increased. The spatiotemporal fusion seismic phase features after frequency domain noise suppression are input into a selective integration unit. A combination of grouped feature remapping and correlation sparse selection is used to generate query features, key features, and value features. Based on feature correlation, effective candidate features for seismic phase picking are selected. The selected effective candidate features are mapped to the same channel time index of the standard DAS spatiotemporal observation tensor according to channel number and then weighted and fused to generate enhanced seismic phase picking features.

8. The spatiotemporal two-dimensional attention fusion method for DAS phase picking according to claim 1, characterized in that: The specific process of outputting the P-wave arrival time and S-wave arrival time based on the enhanced seismic phase picking characteristics and performing ridge constraint correction is as follows: Based on the enhanced phase picking features, output encoding is performed according to the channel time coordinates, and the results are sent to the P-wave output head and S-wave output head for differential analysis to generate P-wave probability sequences and S-wave probability sequences. Based on the P-wave and S-wave probability sequences, candidate arrival times for P-waves and S-waves are located in each channel, and the confidence of the candidate arrival times is adjusted according to the relative arrival times of P-waves and S-waves. Candidate arrival times for P-waves that satisfy cross-channel continuity in adjacent channels are connected to form candidate P-wave ridges, and candidate arrival times for S-waves that satisfy cross-channel continuity in adjacent channels are connected to form candidate S-wave ridges. Ridge constraint correction is applied to the candidate arrival times, and the arrival times of P-waves, S-waves, P-wave confidence, and S-wave confidence of each channel are output.

9. The spatiotemporal two-dimensional attention fusion method for DAS phase picking according to claim 1, characterized in that: The specific process of generating channel-level seismic phase behavior labels and summarizing them to form a seismic phase picking report is as follows: If the P-wave confidence of a channel is lower than the confidence threshold, the channel is marked as having weak P-wave pickup. If the number of S-wave candidate peaks in a channel exceeds the range, or if the S-wave candidate points fall into the tail region, the channel is marked as having multi-peak interference. If the channel quality mask of a certain segment is abnormal and the corresponding phase lacks cross-channel continuity, the segment is marked as having abnormal channel influence. If a segment simultaneously satisfies the first arrival time characteristics, spatial propagation characteristics, and original DAS response characteristics, it is marked as having effective phase propagation. After the channel phase pickup is completed, all channel-level phase behavior labels are summarized. Channel-level label aggregation, propagation ridge consistency verification, and local waveform window verification are used to verify the phase pickup results and obtain a phase pickup report.

10. A spatiotemporal two-dimensional attention fusion system for DAS phase picking, employing the spatiotemporal two-dimensional attention fusion method for DAS phase picking as described in any one of claims 1 to 9, characterized in that, include: The DAS spatiotemporal observation acquisition and benchmarking module is used to acquire raw DAS vibration observation data, and to perform channel direction and time direction reconstruction and amplitude normalization processing, screen candidate seismic phase response zones, and form initial candidate response data. The attention separation parsing module is used to construct a spatial propagation attention map based on the initial candidate response data, generate a temporal arrival attention map, and identify the arrival boundaries of P-waves and S-waves. The dual-gated mutual verification fusion module is used to perform dual-gated mutual verification based on the spatial propagation attention map and the time first arrival attention map, generate spatiotemporal fusion seismic phase features, and enhance the spatiotemporal fusion seismic phase features; The phase ridge constraint output module is used to output the arrival times of P-waves and S-waves based on the enhanced phase picking characteristics, perform ridge constraint correction, generate channel-level phase behavior labels, and summarize them into a phase picking report.

Images