A sea-land differentiated seismic detection positioning method and system for offshore industrial platforms

Abstract

The application discloses a kind of offshore industrial platform-oriented sea-land differentiation seismic detection positioning method and system, method includes: the three-component continuous seismic waveform data of acquisition is divided into marine industrial platform station data stream and land station data stream;According to the noise domain characteristics corresponding to marine industrial platform station data stream and land station data stream respectively, obtain candidate seismic phase set;Candidate seismic phase set is converged into mixed data pool, input space-time seismic correlation algorithm and combine regional crustal velocity model, obtain the correlation seismic phase set of effective seismic event;According to the seismic phase data in correlation seismic phase set, call nonlinear probability earthquake location program to carry out source parameter inversion, and in inversion, include the near-field seismic phase data of marine industrial platform station, output preliminary source location;According to the quality evaluation index of preestablished, preliminary source location is screened, and offshore earthquake catalogue is output.The application successfully overcomes non-stationary noise domain migration.

Description

Technical Field

[0001] This invention belongs to the field of seismological observation and artificial intelligence data processing technology, and in particular relates to a method and system for detecting and locating land-sea differentiated earthquakes for near-shore industrial platforms. Background Technology

[0002] The coastal continental shelf region is not only one of the most densely populated and economically developed areas in the world, but also one of the regions with the highest risk of marine earthquake disasters. Effective monitoring of nearshore earthquakes is crucial for accurately assessing disaster risks and ensuring the safety of coastal engineering projects.

[0003] However, in shallow continental shelf areas, there is still a lack of an economical, efficient, and real-time automated long-term observation and data processing solution. Existing nearshore seismic observation and automated data processing technologies suffer from the following serious shortcomings: 1. Single-sided land-based surveillance networks suffer from severe geometric constraints, and data from near-shore industrial platform stations urgently require support from efficient and automated processing methods. In some nearshore areas, earthquake monitoring relies primarily on stations on the land side, which can easily lead to significant azimuth gaps. Furthermore, the lack of near-epicenter stations creates considerable uncertainty regarding the location and depth of the hypocenter. Deploying seismic stations on nearshore industrial platforms such as offshore wind power substations can supplement marine observations; however, the noise characteristics of data from industrial platform stations differ significantly from those from land-based stations, making it difficult for existing unified processing procedures to simultaneously balance the signal detection rate of marine industrial platform stations with the arrival accuracy of land-based stations. Therefore, a method for automatically processing differentiated data from combined marine and land-based seismic networks is needed.

[0004] 2. The strong resonant noise unique to offshore industrial platforms causes severe "domain shift" in conventional deep learning models: Building an observation network based on near-shore industrial platforms inevitably introduces extremely severe algorithmic challenges. These platforms exhibit significant structural resonances and equipment mechanical noise in specific frequency bands (e.g., 1.2–1.6 Hz), constituting a high-energy, non-stationary noise domain distinctly different from that of land-based stations. Faced with this atypical noise domain, existing automated phase-picking algorithms have revealed serious limitations: Conventional deep learning models pre-trained on land data have extremely poor adaptability: conventional models trained on land environments (such as EQTransformer) exhibit strong conservatism, tending to refuse to recognize low signal-to-noise ratio signals, resulting in a large number of missed detections of minute events; while models based on the U-Net architecture (such as PhaseNet) are very prone to confusing mechanical vibration pulses with the first arrival of earthquake P waves, resulting in an extremely high false detection rate.

[0005] Deep-sea models relying on specific sensors and land-based fine-tuning models are difficult to generalize: Existing noise-resistant models optimized for marine environments (such as PickBlue) often heavily rely on data collaboration from hydrophones (the fourth component) for noise suppression. However, nearshore industrial platforms typically only deploy conventional three-component seismometers, and the lack of crucial hydrophone input leads to a sharp decline in the performance of such deep-sea-specific models. Furthermore, the natural micro-motions in the deep-sea environment (such as ocean currents and water reverberation) differ significantly in spectral characteristics from the strong "structural resonance-mechanical coupling" transient noise of nearshore industrial platforms, making conventional deep-sea models unsuitable for direct application. On the other hand, even models fine-tuned based on regional land data (such as USTC-Pickers) experience a significant performance degradation when facing the extreme mechanical resonance noise of industrial platforms, as the domain shift caused by structural resonance is far greater than the impact of geological structural differences.

[0006] 3. In extreme environments, signal loss and massive false pickups coexist, and there is a lack of systematic process design solutions for such scenarios: Nearshore waters are often covered by thick, loose sedimentary layers, which strongly attenuate high-frequency signals. P-waves from minor earthquakes are often extremely weak when they reach platform stations, easily masked by the platform's continuous mechanical vibrations. Existing conventional models struggle to independently and robustly capture weak S-wave signals when P-waves are completely undetectable. Furthermore, due to the drastically different signal-to-noise ratios between marine and land-based stations, attempting to process the entire marine and land-based network data with a single model is impractical. In this noisy environment, lowering the algorithm's picking threshold to forcibly capture weak signals inevitably introduces a large number of "false picks" triggered by platform mechanical noise. Although existing spatiotemporal correlation algorithms are designed with some noise resistance, there are currently no systematic parameter design principles and process configuration schemes for the specific scenario of nearshore industrial platforms—where the signal-to-noise ratios of marine and land-based stations are extremely unbalanced, and false picks are highly concentrated at industrial platform stations. Without a targeted process design, the association algorithm is prone to misassociations when faced with a dataset containing a large number of false picks, and thus cannot output a reliable earthquake catalog.

[0007] In summary, overcoming the strong structural resonance noise interference from near-shore industrial platforms, eliminating the huge noise domain shift between sea and land stations, and achieving high-precision automated seismic detection and positioning in a combined sea and land environment under extreme conditions where P-waves are easily submerged by noise and false picks surge, are key technical challenges that urgently need to be solved in this field. Summary of the Invention

[0008] To address the aforementioned technical issues, this invention proposes a differentiated seismic detection and location method and system for offshore industrial platforms. Targeting the distinct noise domain characteristics of offshore industrial platform stations compared to land-based stations, a complete automated processing flow is constructed, comprising "station classification and identification—differentiated model allocation—low-threshold tolerance picking—multi-dimensional strong constraint correlation—nonlinear probabilistic location—quality screening."

[0009] To achieve the above objectives, this invention provides a method for detecting and locating land-sea differentiated seismic activity for near-shore industrial platforms, comprising: Based on the deployment environment type represented by the station metadata of each station in the hybrid marine-land network, the collected three-component continuous seismic waveform data is split into marine industrial platform station data stream and land station data stream. Based on the noise domain characteristics of the data streams from the marine industrial platform stations and the land stations, different deep learning phase picking models are invoked to pick P-wave and S-wave phases to obtain a candidate phase set. The candidate phases include at least the station identifier, deployment environment type, phase type, phase arrival time, and picking probability. The candidate seismic phase set is aggregated into a hybrid data pool, input into a spatiotemporal seismic correlation algorithm and combined with a regional crustal velocity model to obtain a set of associated seismic phases for effective seismic events; Based on the arrival time data of the associated seismic phases in the set of associated seismic phases, the nonlinear probabilistic seismic location method is used to invert the earthquake time and the spatial location of the source, and output the preliminary source location and the location uncertainty. Based on the aforementioned location uncertainty, a joint quality screening is performed on the preliminary epicenter location to obtain the final location result, and the final location result is compiled into a nearshore earthquake catalog.

[0010] Optionally, based on the deployment environment type of each station in the hybrid marine-land network, the acquired three-component continuous seismic waveform data can be split into a data stream from marine industrial platform stations and a data stream from land stations, including: Based on the station code, the deployment environment of each station is identified. Data collected by stations deployed on near-shore industrial platforms is divided into the marine industrial platform station data stream, and data collected by conventional seismic observation stations deployed on bedrock or soil on land is divided into the land station data stream.

[0011] Optionally, the step of calling different deep learning phase picking models to perform P-wave and S-wave phase picking based on the noise domain characteristics corresponding to the data streams from the marine industrial platform stations and the land-based stations includes: Based on the data stream from the marine industrial platform stations, a first deep learning model with global context awareness and trained or transferred by real marine observation noise data is invoked to pick up seismic phases. Based on the data stream from the land stations, a second deep learning model trained on earthquake data from China and adapted for the observation area through transfer learning or parameter fitting is invoked to pick up seismic phases.

[0012] Optionally, the first deep learning model is a phase picking model based on the Transformer architecture and equipped with a global self-attention mechanism, which segments the data stream of the marine industrial platform station with a first sliding time window of a preset length of 60 seconds. The second deep learning model is a phase picking model based on a local feature extraction architecture and adapted to regional data. It segments the land station data stream with a second sliding time window of a preset length of 30 seconds. A preset overlap area is set between adjacent sliding time windows, and duplicate candidate phases in the overlap area are merged.

[0013] Optionally, the candidate seismic phase set is aggregated into a hybrid data pool, input into a spatiotemporal seismic correlation algorithm, and combined with a regional crustal velocity model to obtain a set of associated seismic phases for valid seismic events, including: The candidate seismic phase set is aggregated into a hybrid candidate data pool. Four-dimensional spatiotemporal seismic correlation is performed based on theoretical travel time constraints generated by the regional crustal velocity model. Simultaneously, constraints are applied on the total number of correlated seismic phases, the number of P-wave seismic phases, the number of S-wave seismic phases, the number of stations with dual seismic phases, and the travel time residual tolerance to obtain the set of correlated seismic phases of effective earthquake events. The regional crustal velocity model includes multiple velocity nodes or stratigraphic interfaces set along the depth direction, and each velocity node or stratigraphic interface is configured with P-wave velocity parameters and S-wave velocity parameters. During the correlation process, multidimensional strong constraints are applied, requiring that effective earthquake events simultaneously satisfy the following conditions: the total number of correlated seismic phases is not less than the first preset number, the number of P-wave seismic phases is not less than the second preset number, the number of S-wave seismic phases is not less than the third preset number, and at least a fourth preset number of stations simultaneously record P-waves and S-waves, and the difference between the arrival time of the correlated seismic phases and the corresponding theoretical travel time is within the preset travel time residual tolerance.

[0014] Optionally, based on the arrival time data of the associated seismic phases in the set of associated seismic phases, a nonlinear probabilistic seismic location method is used to invert the earthquake occurrence time and spatial location of the epicenter, outputting a preliminary epicenter location and location uncertainty, including: Based on the arrival time data of the associated seismic phases in the set of associated seismic phases, the posterior probability density function of the source parameters is searched in the three-dimensional source space and the range of the time of origin, so that the seismic phases of the marine industrial platform station and the seismic phases of the land station can participate in the localization together, and the maximum likelihood source location, the time of origin, and the source location uncertainty characterized by a preset confidence level are output. Based on the maximum likelihood source location, the time of origin, and the source location uncertainty, the preliminary source location and positioning uncertainty are obtained.

[0015] Optionally, based on the location uncertainty, a joint mass screening is performed on the preliminary source location to obtain the final location result, and the final location result is compiled into a nearshore earthquake catalog, including: The preliminary source location is screened for quality based on at least one of the following indicators: maximum azimuth gap, root mean square travel time residual, semi-major axis of horizontal error ellipsoid, semi-minor axis of horizontal error ellipsoid, and semi-vertical error axis. The location results that meet the quality screening criteria will be output as a high-resolution catalog of nearshore earthquakes, including the time of occurrence, the latitude and longitude of the epicenter, and the focal depth.

[0016] This invention also provides a land-sea differentiated seismic detection and positioning system for near-shore industrial platforms, comprising: a data acquisition and distribution module, a differentiated seismic phase picking module, a candidate seismic phase convergence module, a multidimensional strongly constrained correlation module, a nonlinear probabilistic positioning module, a catalog output module, and a quality assessment and catalog output module; The data acquisition and distribution module is used to distribute the collected three-component continuous seismic waveform data into marine industrial platform station data streams and land station data streams based on the deployment environment type represented by the station metadata of each station in the marine-land hybrid network. The differentiated seismic phase picking module is used to call different deep learning seismic phase picking models to pick P-wave and S-wave seismic phases according to the noise domain characteristics of the data streams of the marine industrial platform stations and the data streams of the land stations, and to obtain a set of candidate seismic phases. The candidate seismic phases include at least the station identifier, deployment environment type, seismic phase type, seismic phase arrival time and picking probability. The candidate seismic phase aggregation module is used to aggregate the candidate seismic phase set into a hybrid data pool, input the spatiotemporal seismic correlation algorithm and combine it with the regional crustal velocity model to obtain the associated seismic phase set of effective seismic events; The nonlinear probabilistic positioning module is used to invert the earthquake occurrence time and spatial location of the source using the nonlinear probabilistic earthquake positioning method based on the earthquake arrival time data in the associated earthquake phase set, and output the preliminary source location and positioning uncertainty; based on the positioning uncertainty, the preliminary source location is subjected to joint quality screening to obtain the final positioning result, and the final positioning result is summarized to form a nearshore earthquake catalog.

[0017] Compared with the prior art, the present invention has the following advantages and technical effects: (1) Improving the detection capability of small earthquake candidate events in high-noise environments: This invention combines a lower candidate phase retention threshold with multi-dimensional joint constraint correlation. While preserving low-probability true phases as much as possible, it suppresses random noise triggering through regional theoretical travel time and multi-station, multi-phase constraints. In continuous waveform data from 2024 to 2026, this invention detected, correlated, and located a total of 140 nearshore earthquake events, while conventional land seismic network catalogs recorded 45 events during the same period. The number of events in the embodiment is approximately three times that of the actual earthquake catalog.

[0018] (2) Improve the geometric constraints of a single-sided land seismic network on nearshore earthquakes: This invention uses quality-controlled seismic phases from marine industrial platform stations and seismic phases from land stations for localization, supplementing the marine side and near-field observations of nearshore events, which can reduce the uncertainty of azimuth gap, epicentral distance from the nearest station, and source location.

[0019] (3) Mitigating the impact of land-sea noise domain differences on seismic phase pickup performance: In the marine industrial platform seismic phase test set of this embodiment, the OBSTransformer, which possesses global context awareness and has been trained with real ocean observation noise enhancement, achieves high overall picking performance; the mean P-wave arrival time residual is 0.09 seconds, and the mean S-wave arrival time residual is 0.004 seconds. Compared with some pre-trained land models in the same test set, its systematic arrival time bias is smaller, which is beneficial for subsequent association and localization. The above results are limited by the dataset and parameter configuration of this embodiment.

[0020] (4) Significantly reduce the cost of marine seismic observation, and empower the safety and earthquake prevention and disaster reduction of major marine engineering projects: This invention, through innovation at the data processing algorithm level, effectively activates the practical value of a new "infrastructure sharing" observation model that deploys conventional seismometers on near-shore industrial platforms such as offshore wind power booster stations. It fundamentally avoids the pain points of traditional seabed seismometers (OBS), such as the inability to transmit data in real time, the need for expensive offshore retrieval operations, susceptibility to damage by trawling nets, and the extremely high cost of laying dedicated submarine optical cables, significantly reducing the construction and operation costs of operational near-shore seismic monitoring. Once deployed, the automated processing flow proposed in this invention can achieve uninterrupted automated processing of continuous waveform data from a hybrid onshore-offshore seismic network, significantly reducing the manpower costs of network operation. Given the explosive growth of the global offshore wind power industry, the method of this invention can be directly replicated and promoted on near-shore industrial platform networks in different regions around the world, possessing extremely high economic reuse value. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a land-sea differentiated seismic detection and location method for near-shore industrial platforms according to an embodiment of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0024] This specific implementation method involves the following technical terms and English abbreviations, the meanings of which are as follows: OBS (Ocean Bottom Seismometer): A traditional seismic observation instrument deployed on the seabed.

[0025] P-wave / S-wave (Primary / Secondary Wave): Seismic longitudinal wave (first arrival compression wave) and seismic transverse wave (second arrival shear wave).

[0026] Phase picking: The process of automatically identifying and calibrating the arrival times of P-waves and S-waves from continuous seismic waveforms.

[0027] Azimuthal gap: In a seismic event and station network, the maximum azimuth angle between two adjacent stations relative to the epicenter. The larger the value, the more uneven the station's coverage of the epicenter.

[0028] OBSTransformer: A deep learning phase picking model based on the Transformer architecture, featuring a global self-attention mechanism, and trained or transferred using real submarine seismic observation events and noisy data.

[0029] USTC-Pickers: A unified deep learning seismic phase picking model set developed by the University of Science and Technology of China. In this embodiment, a region-specific version (USTC-Pickers (Jiangsu)) with fine-tuning for historical data of Jiangsu region is used.

[0030] PyOcto: A high-throughput seismic correlation algorithm based on 4D space-time partitioning.

[0031] NonLinLoc: A nonlinear probabilistic earthquake location method based on octree sampling, which outputs the posterior probability density function (PDF) of the earthquake source location.

[0032] SNR (Signal-to-Noise Ratio): The ratio of seismic signal energy to background noise energy.

[0033] RMS (Root Mean Square): The root mean square of the travel time residuals measures the overall goodness of fit between the observed and theoretical arrival times of seismic phases.

[0034] GNSS (Global Navigation Satellite System): A global satellite navigation system used for high-precision time synchronization of satellite stations.

[0035] MEMS (Micro-Electro-Mechanical System): In this context, it refers to a high-precision three-component strong-motion seismometer based on microelectromechanical technology.

[0036] VPN (Virtual Private Network): In this context, it refers to an encrypted data transmission channel established through a carrier's 4G wireless network.

[0037] Location uncertainty: The degree of uncertainty of the source location characterized by the posterior probability distribution or confidence region of the nonlinear probabilistic location results, including horizontal location uncertainty and depth location uncertainty.

[0038] In view of the technical bottlenecks faced by existing technologies in automated monitoring and data processing of nearshore earthquakes, the specific technical problems that this embodiment aims to solve are as follows: (1) Solving the problem of severe "domain shift" and model generalization failure of conventional models in the high-noise environment of industrial platforms. For the strong structural resonance and mechanical transient noise encountered by high-sensitivity seismometers deployed on near-shore industrial platforms, conventional deep learning models are prone to missing weak earthquakes or falsely detecting mechanical impulses. Furthermore, existing deep-sea-specific models are unsuitable due to a lack of hydrophone data. The purpose of this invention is to provide a sea-land differentiated deep learning phase picking strategy, abandoning the traditional paradigm of a single model for the entire network. Based on the distinct noise domain characteristics of marine industrial platform stations and land stations, a deep learning phase picking model adapted to their respective noise characteristics is matched. While effectively overcoming platform mechanical noise interference and successfully extracting weak seismic signals, the picking accuracy advantage of land stations is fully preserved, achieving overall optimal performance for the entire sea-land seismic network.

[0039] (2) Solving the problem of automated correlation failure caused by the coexistence of missing P-wave signals and massive false picking in extremely noisy environments. The strong attenuation of high-frequency P-wave signals by the thick, loose sedimentary layer in nearshore areas, and the inevitable false triggering by massive platform mechanical noise after lowering the picking threshold, lead to the inherent defects in existing conventional seismic correlation algorithms, which are prone to miscorrelation or even processing failure. The purpose of this invention is to establish a collaborative process mechanism of "differentiated model allocation—low-threshold tolerant picking—multi-dimensional strongly constrained spatiotemporal correlation." On the one hand, by introducing a noise-resistant model with global context awareness, it achieves independent and robust capture of weak S-wave signals even when P-waves are completely submerged in noise; on the other hand, through a high-precision spatiotemporal partitioning correlation algorithm and stringent multi-station, multi-phase constraints, it robustly locks and isolates real small seismic events from massive false picking data, ensuring the high reliability of the automated processing flow.

[0040] (3) Addressing the problem that near-shore industrial platform station data cannot be effectively utilized by existing automated processes, resulting in consistently low accuracy of marine seismic positioning. This invention addresses the technical shortcomings of existing automated processing methods, which cannot fully utilize marine station data due to their inability to adapt to the noise domain of industrial platforms, leading to insufficient constraints on the location of marine earthquake sources, particularly their depth. The purpose of this invention is to effectively integrate near-shore industrial platform observation data into the network through the aforementioned highly robust automated processing flow, providing sufficient near-field phase constraints, significantly compressing the source error space, improving source depth positioning accuracy, and ultimately providing an economical, efficient, and real-time monitoring-enabled integrated land-sea automatic positioning solution, offering a high-quality seismic catalog for the detailed analysis of near-shore seismogenic structures.

[0041] Based on this, this embodiment proposes a land-sea differentiated seismic detection and location method for near-shore industrial platforms, such as... Figure 1 As shown, the specific steps include: Based on the deployment environment type represented by the station metadata of each station in the hybrid marine-land network, the collected three-component continuous seismic waveform data is split into marine industrial platform station data stream and land station data stream. Based on the noise domain characteristics of the data streams from the marine industrial platform stations and the land stations, different deep learning phase picking models are invoked to pick P-wave and S-wave phases to obtain a candidate phase set. The candidate phases include at least the station identifier, deployment environment type, phase type, phase arrival time, and picking probability. The candidate seismic phase set is aggregated into a hybrid data pool, input into a spatiotemporal seismic correlation algorithm and combined with a regional crustal velocity model to obtain a set of associated seismic phases for effective seismic events; Based on the arrival time data of the associated seismic phases in the set of associated seismic phases, the nonlinear probabilistic seismic location method is used to invert the earthquake time and the spatial location of the source, and output the preliminary source location and the location uncertainty. Based on the aforementioned location uncertainty, a joint quality screening is performed on the preliminary epicenter location to obtain the final location result, and the final location result is compiled into a nearshore earthquake catalog.

[0042] Specifically, S1: Input and splitting steps for continuous waveform data from the hybrid marine-land network. The system receives or reads three-component continuous seismic waveform data collected by the hybrid marine-land network in real time or offline. Based on the physical environment of the station deployment, the data is split into two streams: "Marine Industrial Platform Station Data Stream" and "Land Station Data Stream," which are then sent to subsequent differentiated processing branches.

[0043] S21: Data stream from offshore industrial platform stations. This stream is generated by the diversion in step S1 and includes three-component continuous waveform data from all stations deployed on nearshore industrial platforms (including but not limited to offshore wind power booster stations, drilling platforms, offshore oil platforms, and other facilities facing strong resonance environments). It then proceeds to step S23 for phase picking.

[0044] S22: Land station data stream. Generated from step S1, it contains three-component continuous waveform data from all conventional seismic observation stations deployed on bedrock or soil, and proceeds to step S24 for phase picking.

[0045] S23: Differentiated Seismic Phase Picking Module for Marine Industrial Platform Stations. For the data stream of S21, a deep learning model based on the Transformer architecture and trained with real ocean strong noise enhancement is used to pick P-wave and S-wave phases (OBSTransformer model is preferred), and outputs a set of candidate phases for marine stations.

[0046] S24: Differentiated phase picking module for land stations. For the data stream of S22, a mature deep learning model that has been fine-tuned locally based on historical land data of the observation area is used to pick P-wave and S-wave phases (preferably the USTC-Pickers regional fine-tuning model), and outputs a set of candidate phases for land stations.

[0047] S3: Module for merging land and sea seismic phase picking data and using a low threshold for filtering. This module merges the two candidate seismic phase sets output from steps S23 and S24, and filters them uniformly using an extremely low picking probability threshold (preferably 0.1). The resulting pool of mixed candidate seismic phase data, containing both genuine weak seismic phases and false picking due to mechanical noise, is then fed into step S4. The technical purpose of setting a low probability threshold is to maximize the preservation of weak, genuine seismic signals masked by high-energy noise. This, together with the multi-dimensional strong constraint correlation in step S4, constitutes a "wide entry, strict exit" collaborative error correction mechanism.

[0048] S4: Seismic correlation module based on 4D spatiotemporal partitioning and multidimensional strong constraints. The hybrid data pool output from step S3 is input into a high-throughput correlation algorithm based on four-dimensional spatiotemporal partitioning (preferably PyOcto algorithm). Travel time constraints are provided by combining a regional one-dimensional crustal velocity model, while extremely stringent multidimensional strong constraints are imposed (requiring a single event to meet: total number of seismic phases ≥ 8, P-wave number ≥ 4, S-wave number ≥ 2, number of dual-phase stations ≥ 2). Through these strong constraints, real seismic events are robustly identified from a massive amount of false picking, and spurious signals are removed, outputting a set of valid seismic phases that have passed correlation verification.

[0049] S5: Joint land-sea nonlinear probabilistic earthquake location module. The effective phase arrival time data output from step S4 is combined with the regional one-dimensional crustal velocity model and input into the nonlinear probabilistic earthquake location program (preferably the NonLinLoc method). The program calculates the spatial coordinates of the seismic source and the time of origin through a three-dimensional spatial grid search, making full use of the near-field constraints of the marine stations, and outputs the time of origin, preliminary source location, and source location uncertainty.

[0050] S6: Location Quality Assessment and High-Resolution Earthquake Catalog Output Module. This module performs parameter statistics and filtering on the location results output from step S5. Evaluation indicators include the maximum azimuth gap, root mean square travel time residual (RMS), and source location uncertainty. Based on this, further physical accuracy verification can be performed by combining regional wavefield propagation patterns (such as the time-distance curve and the wave velocity ratio from the datum). The final output is a highly reliable, high-resolution nearshore earthquake catalog containing the earthquake's origin time, latitude and longitude, source depth, and location quality parameters.

[0051] Furthermore, based on the deployment environment type of each station in the hybrid marine-land network, the collected three-component continuous seismic waveform data is split into a data stream from marine industrial platform stations and a data stream from land stations, including: Based on the station code, the deployment environment of each station is identified. Data collected by stations deployed on near-shore industrial platforms is divided into the marine industrial platform station data stream, and data collected by conventional seismic observation stations deployed on bedrock or soil on land is divided into the land station data stream.

[0052] Specifically, three-component continuous seismic waveform data were acquired from a mixed land-sea seismic network. The waveform data stream was divided into two categories according to the station codes: (1) Marine industrial platform data: refers to seismic observation stations deployed on marine engineering platforms (but not limited to offshore wind power booster stations, drilling platforms, offshore oil platforms, and other near-shore engineering platforms). These stations are affected by the inherent dynamic characteristics of the steel structure pile foundation of the platform, and have a significant structural resonance amplification effect in a specific frequency band (e.g., 1–2Hz, which is related to the size and structure of the platform itself). The background noise exhibits high-energy, non-stationary characteristics, which is fundamentally different from the conventional noise domain.

[0053] (2) Land station data: refers to conventional seismic observation stations deployed on the bedrock or soil of land. The background noise level of such stations is relatively stable, and the signal-to-noise ratio distribution of each frequency band conforms to the conventional seismic observation law.

[0054] This is a prerequisite for achieving differentiated processing. By separating the land and sea data, two fundamentally different noise domains are clearly distinguished, avoiding the performance failure of a single data processing pipeline when faced with extreme signal-to-noise ratio differences between land and sea, and providing structured data input for subsequent differentiated model invocation.

[0055] Furthermore, the step of calling different deep learning phase picking models to perform P-wave and S-wave phase picking based on the noise domain characteristics corresponding to the data streams from the marine industrial platform stations and the land-based stations includes: Based on the data stream from the marine industrial platform stations, a first deep learning model with global context awareness and trained or transferred by real marine observation noise data is invoked to pick up seismic phases. Based on the data stream from the land stations, a second deep learning model trained on earthquake data from China and adapted for the observation area through transfer learning or parameter fitting is invoked to pick up seismic phases.

[0056] The first deep learning model is a phase picking model based on the Transformer architecture and equipped with a global self-attention mechanism. It segments the data stream of the marine industrial platform station with a first sliding time window of a preset length of 60 seconds. The second deep learning model is a phase picking model based on a local feature extraction architecture and adapted to regional data. It segments the land station data stream with a second sliding time window of a preset length of 30 seconds. A preset overlap area is set between adjacent sliding time windows, and duplicate candidate phases in the overlap area are merged.

[0057] Specifically, for the two sets of data after splitting, deep learning models with different architectures and training domains with different noise levels are used to pick up P-wave and S-wave phases, and a picking probability threshold lower than the conventional standard is uniformly set. This includes: (1) Seismic phase pickup of offshore industrial platforms: Technical approach: A deep learning model (preferably the OBSTransformer model) based on the global attention mechanism and trained with real ocean noise enhancement is used for processing. The picking probability threshold is set to a low tolerance threshold (preferably 0.1, which can be adjusted within the range of 0.05 to 0.3 according to the actual background noise level of the station).

[0058] Operating principle: Vibrations of industrial platform structures can cause some mechanical pulses to resemble the initial arrival of an earthquake in terms of local waveform morphology. Models with long-term global context awareness can synthesize the three-component waveform characteristics over a longer time range. In the test set of this embodiment, the model can still provide candidate results for subsequent correlation verification for some low signal-to-noise ratio seismic phases, and reduces systematic arrival time bias compared to some land-based models; its applicability still needs to be evaluated in conjunction with specific platform noise and test data.

[0059] (2) Seismic phase pickup at land stations: A mature regional model (preferably a version of the USTC-Pickers series fine-tuned for the corresponding region) based on a local feature extraction architecture and adapted to the observation area is used for processing. A low tolerance threshold (preferably 0.1) is set for the picking probability threshold. Taking full advantage of the low background noise and high signal-to-noise ratio of land stations, a mature model learned through transfer learning from regional measured data ensures high-precision and high-sensitivity picking of land seismic data, guarantees the continuity and integrity of the entire network monitoring, and provides high-quality land station seismic phase constraints for subsequent correlation steps.

[0060] The overall purpose of this step is to maximize the recall of weak, real seismic signals masked by platform noise by actively lowering the picking threshold ("tolerant picking"), ensuring that extremely weak events are not missed. Although this operation introduces a large number of false picks triggered by mechanical noise, it provides a complete candidate data pool containing real, extremely weak signals for subsequent strong-constraint correlation, forming a "broad entry, strict exit" collaborative mechanism with step three.

[0061] Furthermore, the candidate seismic phase set is aggregated into a hybrid data pool, input into a spatiotemporal seismic correlation algorithm, and combined with a regional crustal velocity model to obtain a set of associated seismic phases for valid seismic events, including: The candidate seismic phase set is aggregated into a hybrid candidate data pool. Four-dimensional spatiotemporal seismic correlation is performed based on theoretical travel time constraints generated by the regional crustal velocity model. Simultaneously, constraints are applied on the total number of correlated seismic phases, the number of P-wave seismic phases, the number of S-wave seismic phases, the number of stations with dual seismic phases, and the travel time residual tolerance to obtain the set of correlated seismic phases of effective earthquake events. The regional crustal velocity model includes multiple velocity nodes or stratigraphic interfaces set along the depth direction, and each velocity node or stratigraphic interface is configured with P-wave velocity parameters and S-wave velocity parameters. During the correlation process, multidimensional strong constraints are applied, requiring that effective earthquake events simultaneously satisfy the following conditions: the total number of correlated seismic phases is not less than the first preset number, the number of P-wave seismic phases is not less than the second preset number, the number of S-wave seismic phases is not less than the third preset number, and at least a fourth preset number of stations simultaneously record P-waves and S-waves, and the difference between the arrival time of the correlated seismic phases and the corresponding theoretical travel time is within the preset travel time residual tolerance.

[0062] The one-dimensional crustal velocity model for the region can be implemented using either a layered constant velocity form or a depth node interpolation form. Each stratum or depth node includes at least depth, P-wave velocity, and S-wave velocity parameters. Before association, theoretical travel times for P-waves and S-waves are pre-calculated or calculated in real-time based on station coordinates and the spatial range of candidate seismic sources, for use in determining the spatiotemporal consistency of candidate seismic phases.

[0063] Furthermore, based on the arrival time data of the associated seismic phases in the set of associated seismic phases, the nonlinear probabilistic seismic location method is used to invert the earthquake occurrence time and spatial location of the epicenter, outputting the preliminary epicenter location and location uncertainty, including: Based on the arrival time data of the associated seismic phases in the set of associated seismic phases, the posterior probability density function of the source parameters is searched in the three-dimensional source space and the range of the time of origin, so that the seismic phases of the marine industrial platform station and the seismic phases of the land station can participate in the localization together, and the maximum likelihood source location, the time of origin, and the source location uncertainty characterized by a preset confidence level are output. Based on the maximum likelihood source location, the time of origin, and the source location uncertainty, the preliminary source location and positioning uncertainty are obtained.

[0064] Specifically, the output set of mixed seismic phases, containing both real and numerous false triggers, is input into a high-throughput correlation algorithm (preferably PyOcto) based on four-dimensional space-time partitioning for spatiotemporal clustering. The correlation process incorporates a one-dimensional crustal velocity model of the observation area to provide travel time constraints. Simultaneously, stringent multi-dimensional correlation constraints are set; a valid seismic event must simultaneously satisfy all of the following conditions: (1) The total number of associated seismic phases shall not be less than Nmin (preferably 8, which can be adjusted according to the actual situation such as station data and station spacing). (2) The number of P waves contained therein is not less than Np (preferably 4, which can be adjusted according to the actual situation such as station data and station spacing). (3) The number of S waves included shall not be less than Ns (preferably 2, which can be adjusted according to the actual situation such as station data and station spacing). (4) At least M stations simultaneously record P-waves and S-waves (preferably 2 stations, which can be adjusted according to station data, station spacing and other actual conditions, and used in conjunction with travel time residual tolerance).

[0065] This step is the core of solving the problem of "association failure caused by false picking". The efficient 4D spatiotemporal partitioning algorithm ensures computational efficiency under massive mixed seismic phase data; the stringent multi-station and multi-phase joint constraints utilize the physical inevitability of real seismic events satisfying regional travel time patterns at multiple stations, which is fundamentally distinguished from the spatiotemporal irregularity of false picking triggered by random noise. Thus, it robustly locks and isolates real small seismic events from the massive false picking data pool, ensuring the high reliability of the association results.

[0066] The arrival times of successfully correlated seismic events are combined with one-dimensional or three-dimensional crustal velocity models of the observation area and input into a nonlinear probabilistic seismic location program (preferably the NonLinLoc nonlinear probabilistic location method) for source parameter inversion. NonLinLoc uses an octree sampling algorithm to search for the posterior probability density function (PDF) of the source location in three-dimensional space, outputting the maximum likelihood epicenter location, time of occurrence, and its three-dimensional error ellipsoid parameters, including the semi-major axis, semi-minor axis, and vertical error semi-axis of the horizontal error ellipse. During location, all valid seismic phases from marine industrial platform stations and land-based stations that have passed quality control are preferentially included.

[0067] By incorporating near-field or marine-side effective seismic phases from marine industrial platform stations into the localization process along with terrestrial seismic phases, the azimuth coverage of a single-sided terrestrial seismic network can be supplemented, reducing the dispersion of the posterior probability distribution of the source location, and particularly helping to improve the source depth constraint.

[0068] Furthermore, based on the aforementioned location uncertainty, a joint mass screening is performed on the preliminary epicenter location to obtain the final location result, and the final location result is compiled into a near-shore earthquake catalog, including: The preliminary source location is screened for quality based on at least one of the following indicators: maximum azimuth gap, root mean square travel time residual, semi-major axis of horizontal error ellipsoid, semi-minor axis of horizontal error ellipsoid, and semi-vertical error axis. The location results that meet the quality screening criteria will be output as a high-resolution catalog of nearshore earthquakes, including the time of occurrence, the latitude and longitude of the epicenter, and the focal depth.

[0069] Specifically, the positioning results output in step S5 are statistically analyzed for indicators such as maximum azimuth gap, root mean square of travel time residual, horizontal position uncertainty, and depth position uncertainty, and then jointly screened according to preset thresholds. By standardizing quality assessment indicators, the scientific reliability of the final earthquake catalog is ensured, providing a high-quality data foundation for subsequent scientific research such as nearshore seismogenic structure analysis and seismic hazard assessment.

[0070] In other embodiments, the seismic phase picking model, seismic correlation algorithm, and seismic source location algorithm can be replaced by algorithms with corresponding functions. However, the alternatives should be used in conjunction with the station environment diversion, candidate seismic phase retention, regional theoretical travel time constraints, multi-dimensional joint correlation, and sea-land joint location steps of this invention. Alternatives for the marine station seismic phase picking model: The OBSTransformer model in step two can be replaced with other deep learning models that meet the following conditions: possessing a global attention mechanism or equivalent long-term temporal context awareness capability, and having undergone noise-resistant seismic phase picking model training or transfer learning on a real industrial platform or marine environment noise enhancement platform, including but not limited to noise-resistant models using the Transformer architecture, or new specialized models developed in the future.

[0071] Alternatives for land station models: The USTC-Pickers model in step two can be replaced with a deep learning phase picking model that has been trained or fine-tuned with localized data from the observation area and performs well on land station data, including but not limited to mature models such as PhaseNet (DiTing).

[0072] Alternative correlation algorithm: The alternative seismic correlation algorithm should be able to determine the spatiotemporal consistency of candidate P-wave and S-wave phases from multiple stations under the theoretical travel time constraint of the regional crustal velocity model, and implement multi-dimensional joint constraints.

[0073] Alternative positioning algorithms: Alternative positioning algorithms should be able to estimate the source location, time of origin, and positioning uncertainty using the phase arrival time and uncertainty of the mixed land and sea seismic network.

[0074] The following is a detailed description of this embodiment: This embodiment takes the construction of a hybrid monitoring network in nearshore waters as an example to provide the physical basis for the implementation of the method.

[0075] 1. Selection and installation of offshore industrial platform stations: Eight offshore wind power booster stations located in the southern continental shelf area of ​​Sea A were selected as deployment platforms (station codes HYB01, HYB03 to HYB08 and HYS02, located 30–100 km from the shore). To reduce upper-level sway interference, the seismometers were rigidly connected to a platform steel plate next to the bottom load-bearing column via a specially designed steel base, and equipped with a special protective cover suitable for the high-salt and high-humidity marine environment.

[0076] Sensor configuration: HYB series stations are preferably equipped with wideband seismometers with a bandwidth of 60s–50Hz and a sampling rate of 100Hz; HYS02 stations are preferably equipped with short-period seismometers with a bandwidth of 2s–50Hz and a sampling rate of 100Hz. All stations are preferably equipped with high-precision MEMS three-component strong-motion seismometers (sampling rate 200Hz) at the same location, and high-precision time synchronization (accuracy better than 1 millisecond) is achieved through a top deck GNSS antenna. The system is powered by industrial-grade AC power converted to DC through voltage regulation.

[0077] This embodiment does not limit the specific model of the sensor. Any wideband or short-period seismometer with the corresponding bandwidth and sampling rate can be used in this method; any industrial power supply solution that can provide a stable DC power supply can be used as a substitute.

[0078] 2. Real-time data transmission system architecture: A hybrid dual-link transmission scheme is preferred: the primary link utilizes existing submarine optical cables on near-shore industrial platforms to establish a physical leased line to access the data center (bandwidth ≥ 10Mbps, latency < 20ms); the backup link utilizes an industrial-grade router to establish a VPN tunnel to the carrier's wireless network. Automatic switching occurs when the primary link is interrupted, ensuring real-time and continuous data transmission. This invention does not limit the specific implementation of the transmission link; any communication scheme that can guarantee stable real-time data transmission is applicable.

[0079] 3. Land station network: Conventional land seismic stations (preferably deployed in bedrock or stable soil layers, equipped with broadband seismometers) within a 100-kilometer radius of the coastline are selected to participate in the joint network. The number of stations can be determined according to the actual scale of the regional network; in this embodiment, there are 43 stations.

[0080] Based on the aforementioned hardware network, this embodiment provides a general execution flow for automated processing methods.

[0081] Step S1: Acquisition and logic splitting of multi-source continuous waveform data: Real-time reading of three-component continuous waveform data from the hybrid land-sea geostation network. Based on the physical environment of the station deployment, the data stream is explicitly divided into two paths, each sent to a different processing branch to avoid performance failure of a single processing pipeline due to extreme signal-to-noise ratio differences between land and sea. S21 (Offshore Industrial Platform Station Data): This includes station data deployed on nearshore industrial platforms (including but not limited to offshore wind power booster stations, drilling platforms, offshore oil platforms, etc.). This type of data is affected by the dynamic characteristics of steel structure pile foundations, and exhibits a strong structural resonance amplification effect in a specific frequency band (1.2–1.6Hz in this embodiment), with background noise exhibiting high-energy and non-stationary characteristics.

[0082] S22 (Land Station Data): Includes data from conventional land stations, with relatively stable noise levels and a signal-to-noise ratio distribution that conforms to conventional seismological observation patterns.

[0083] Step S2: Deep learning phase picking based on land-sea differences and low thresholds: Marine station processing branch (S23): This branch feeds the S21 data into the OBSTransformer model for P / S wave picking, processes the three-component data with a 60-second sliding time window, and outputs P-wave and S-wave probability sequences. This model utilizes a global context self-attention mechanism and is trained or transfer-learned using real submarine seismic observation events and noisy data. In this embodiment's platform test set, it demonstrates good overall picking performance.

[0084] Land station processing branch (S24): The S22 data is fed into the USTC-Pickers (Jiangsu) model (based on U-Net architecture), and the three-component data is processed with a 30-second sliding time window. The P-wave and S-wave probability sequences are output, focusing on the high-frequency abrupt change features near the first arrival, giving full play to the accuracy advantage of mature regional models in high signal-to-noise ratio land environments.

[0085] Tolerant Picking Settings (S3 Module): The picking probability threshold for both branches is uniformly set to 0.1 (which can be adjusted within the range of 0.05 to 0.3 according to the actual noise level). This "low-threshold tolerant picking" aims to maximize the recall of weak real seismic signals masked by noise. The massive number of false triggers generated are identified and eliminated by the strong constraint correlation mechanism in the subsequent step S4. Together, they constitute a "lenient entry and strict exit" collaborative error correction mechanism.

[0086] Step S4: Convergence of candidate seismic phases and spatiotemporal correlation of multidimensional joint constraints.

[0087] The seismic phase set output from the dual branches in step S2 is pooled into a hybrid candidate pool and input into the PyOcto correlation algorithm. Travel time constraints are provided by combining the regional one-dimensional crustal velocity model, imposing stringent strong topological constraints in the 4D spatiotemporal partitioning clustering: a valid event must simultaneously satisfy the following conditions: total number of seismic phases ≥ 8, P-wave ≥ 4, S-wave ≥ 2, and at least two stations simultaneously recording both P-wave and S-wave. These constraint values ​​can be adjusted according to the actual station distribution density and inter-station spacing, while meeting preset travel time residual tolerances.

[0088] This step leverages the inherent difference between the real earthquake travel time conforming to regional physical laws and the spatiotemporal irregularity of false picking triggered by random mechanical resonance of industrial platforms, thus robustly locking in real earthquake events.

[0089] Step S5: Joint land-sea nonlinear probabilistic seismic location.

[0090] Input the arrival times and uncertainties of the seismic phases of the marine industrial platform stations and land stations that were successfully associated and passed quality control in step S4 into the NonLinLoc program. Calculate the posterior probability density function of the source parameters through octree sampling to obtain the earthquake time, maximum likelihood source location, and horizontal and depth positioning uncertainties.

[0091] Step S6: Joint assessment of positioning quality and output of the catalog. Based on the results of Step S5, at least the maximum azimuth gap, root mean square of travel time residuals, horizontal position uncertainty, and depth position uncertainty are statistically analyzed. Joint screening is performed according to preset thresholds to obtain the final positioning results, which are then compiled into a nearshore earthquake catalog.

[0092] To verify the superiority of the above technical solution, this embodiment provides evaluation results based on real data from the above-mentioned network.

[0093] Verification 1: Scientific validity of the differentiated allocation strategy (comparative test): Using 77 high-quality marine seismic phases independently cross-validated by three senior experts as a benchmark (43 P-waves and 34 S-waves; tolerance of 0.5s for P-waves and 1.0s for S-waves), six mainstream deep learning models were evaluated under a unified testing framework to demonstrate the necessity of the differentiated selection strategy of this invention. Comparative model: EQTransformer (Land Baseline Model): Performed extremely conservatively, with P-wave / S-wave recall rates of only 55.8% / 50.0%, missing a large number of weak events, and a P-wave F1 score of only 0.69; PhaseNet (Land High Sensitivity Model): Extremely high false detection rate (false P-wave pickup FP=31), easily misidentifies mechanical vibration pulses as the first arrival of P-wave, P-wave F1 value is only 0.64; PickBlue (Deep-sea Optimized Model): It is highly dependent on the fourth component input of the hydrophone. Its performance drops sharply in industrial platform scenarios with only three-component seismometers, with the lowest P-wave F1 value being only 0.61. USTC-Pickers (Jiangsu) (Regional Land Fine-tuning Model): Under strong resonance noise, the P-wave F1 value plummeted to 0.58, proving that the domain migration caused by structural resonance has a much greater constraint on model performance than geological structural differences, and the conventional fine-tuning model failed to generalize.

[0094] Preferred model of the present invention: The OBSTransformer model selected for the marine branch of this invention achieved optimal overall performance: P-wave precision 0.82, recall 0.86, and F1 score 0.84; S-wave recall 0.88 and F1 score 0.69. Residual analysis further verified its positioning applicability: the mean residual for P-wave arrival time was only 0.09s, and the mean residual for S-wave was close to zero (0.004s), demonstrating excellent unbiasedness; while in the comparative example, PhaseNet showed a significant positive systematic bias in S-wave arrival time (mean 0.18s), directly affecting subsequent positioning accuracy. These results indicate that, under this test set and parameter configuration, OBSTransformer is more suitable as the preferred model for the processing branch of marine industrial platform stations, but it does not constitute an absolute limitation on the performance of other datasets.

[0095] Verification 2: Overall efficiency and reliability of the automated testing process (actual testing from August 2024 to April 2026): The complete process of this embodiment was applied to continuous waveform data from August 2024 to April 2026, detecting and locating a total of 140 seismic events. Compared to the 45 events (ML1.8-ML4.0) recorded in the conventional seismic network catalog during the same period, this invention newly detected 95 earthquakes (ML1.6-ML2.6), and the number of earthquakes detected in the catalog is approximately 3.1 times the number of earthquakes recorded by the seismic network, with the smallest detected magnitude reaching ML1.6.

[0096] The travel times of all 140 events were compared with historical travel time data for the region over the past 23 years (1998–2021). All data points closely matched the regional P / S wave travel time curves, with no obvious outliers. The marine station data points primarily filled the near-field observation gaps at epicentral distances of 0–50 km. The correlation coefficient with the Datu was 0.972, and the wave velocity ratio was 1.701, highly consistent with the regional geological background. Dual physical verification confirmed that all detected earthquakes were genuine tectonic earthquakes, not spurious correlations, strongly demonstrating the effectiveness of the "low-threshold tolerance picking—multidimensional strong constraint correlation" synergistic mechanism in removing spurious signals and identifying extremely weak genuine earthquakes in extremely noisy environments.

[0097] Verification 3: The method of this invention enables a systematic leap in positioning accuracy after incorporation of near-field seismic phases in the ocean (quantitative assessment of 20 marine events). Because the differentiated processing method of this invention successfully solves the domain offset problem in seismic phase picking for industrial platforms (see Verification 1), high-quality near-field seismic phases from marine stations can be truly extracted and integrated into the positioning calculation (see Verification 2). If a conventional single model is used to process marine station data, near-field constraints cannot be effectively established due to a large number of missed and false seismic phases, and the following improvement in positioning accuracy cannot be achieved. Based on 20 high-quality near-shore earthquakes captured during system operation, the performance improvement is quantified by comparing the positioning parameters of the "invention's combined land-sea processing scheme" and "using only a single-sided land network," as shown in Table 1: Table 1 The significant reduction of 34.0% in focal depth error is particularly noteworthy. For nearshore waters covered by thick Quaternary loose sedimentary layers, high-precision depth constraints are a key criterion for distinguishing between shallow sedimentary microseisms and deep basement fault activity, possessing irreplaceable scientific value for identifying nearshore seismogenic structures. Taking a typical earthquake in the Rudong sea area as an example, after introducing seismic phases from marine stations, the maximum azimuth gap decreased sharply from 223.3° to 104.5°, and the scatter plot of the focal location probability density function converged significantly. This event successfully transformed from an "external event" controlled by a single-sided station network into a well-constrained "internal event," achieving a fundamental leap in nearshore earthquake location accuracy.

[0098] Under the given data and parameter configurations in this embodiment, the detection catalog generated by this method contains 140 seismic events, approximately 3.1 times the 45 events in a conventional land seismic network catalog during the same period. In a comparison of joint land-sea location data for 20 nearshore earthquakes with location data from land stations alone, the depth location uncertainty index decreased by an average of 34.0%. These results demonstrate that the present invention can improve the completeness of the nearshore weak earthquake catalog and enhance the azimuth coverage and source location constraints of nearshore events.

[0099] This embodiment also provides a land-sea differentiated seismic detection and positioning system for near-shore industrial platforms, including: a data acquisition and distribution module, a differentiated seismic phase picking module, a candidate seismic phase aggregation module, a multi-dimensional strongly constrained correlation module, a nonlinear probabilistic positioning module, a catalog output module, and a quality assessment and catalog output module. The data acquisition and distribution module is used to distribute the collected three-component continuous seismic waveform data into marine industrial platform station data streams and land station data streams based on the deployment environment type represented by the station metadata of each station in the marine-land hybrid network. The differentiated seismic phase picking module is used to call different deep learning seismic phase picking models to pick P-wave and S-wave seismic phases according to the noise domain characteristics of the data streams of the marine industrial platform stations and the data streams of the land stations, and to obtain a set of candidate seismic phases. The candidate seismic phases include at least the station identifier, deployment environment type, seismic phase type, seismic phase arrival time and picking probability. The candidate seismic phase aggregation module is used to aggregate the candidate seismic phase set into a hybrid data pool, input the spatiotemporal seismic correlation algorithm and combine it with the regional crustal velocity model to obtain the associated seismic phase set of effective seismic events; The nonlinear probabilistic positioning module is used to invert the earthquake occurrence time and spatial location of the source using the nonlinear probabilistic earthquake positioning method based on the earthquake arrival time data in the associated earthquake phase set, and output the preliminary source location and positioning uncertainty; based on the positioning uncertainty, the preliminary source location is subjected to joint quality screening to obtain the final positioning result, and the final positioning result is summarized to form a nearshore earthquake catalog.

[0100] Specifically, the data acquisition and distribution module is used to acquire three-component continuous waveform data collected by the mixed land and sea network, and divide the waveform data into marine industrial platform station data stream and land station data stream according to the physical environment type of the station deployment.

[0101] The differentiated seismic phase picking module calls the first deep learning model trained with real ocean observation noise data or transfer learning, and the second deep learning model trained with Chinese regional earthquake data and adapted to the observation area. It outputs P-wave and S-wave probability sequences with a sliding time window and extracts candidate seismic phases according to the candidate seismic phase retention thresholds set by each module.

[0102] The candidate seismic phase aggregation module establishes a mixed candidate data pool according to a unified field format. The multi-dimensional joint constraint correlation module uses theoretical travel times generated by a four-dimensional spatiotemporal partitioning algorithm and a regional crustal velocity model to apply constraints on the total number of seismic phases, P-wave number, S-wave number, number of stations with dual seismic phases, and travel time residual tolerance, extracting a set of effective correlated seismic phases. The nonlinear probabilistic positioning module takes the arrival times and uncertainties of seismic phases from marine industrial platform stations and land stations (through quality control) as input to the positioning program, and outputs the earthquake time, maximum likelihood source location, and positioning uncertainty. The quality assessment and catalog output module performs joint screening based on multiple quality indicators, outputting a catalog of nearshore earthquakes containing the earthquake time, epicenter latitude and longitude, source depth, number of correlated stations, number of correlated seismic phases, and positioning quality parameters.

[0103] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting and locating land-sea differentiated seismic activity for near-shore industrial platforms, characterized in that, include: Based on the deployment environment type represented by the station metadata of each station in the hybrid marine-land network, the collected three-component continuous seismic waveform data is split into marine industrial platform station data stream and land station data stream. Based on the noise domain characteristics of the data streams from the marine industrial platform stations and the land stations, different deep learning phase picking models are invoked to pick P-wave and S-wave phases to obtain a candidate phase set. The candidate phases include at least the station identifier, deployment environment type, phase type, phase arrival time, and picking probability. The candidate seismic phase set is aggregated into a hybrid data pool, input into a spatiotemporal seismic correlation algorithm and combined with a regional crustal velocity model to obtain a set of associated seismic phases for effective seismic events; Based on the arrival time data of the associated seismic phases in the set of associated seismic phases, the nonlinear probabilistic seismic location method is used to invert the earthquake time and the spatial location of the source, and output the preliminary source location and the location uncertainty. Based on the aforementioned location uncertainty, a joint quality screening is performed on the preliminary epicenter location to obtain the final location result, and the final location result is compiled into a nearshore earthquake catalog.

2. The method for detecting and locating differential seismic activity between land and sea for near-shore industrial platforms according to claim 1, characterized in that, Based on the deployment environment type of each station in the hybrid marine-land seismic network, the collected three-component continuous seismic waveform data is split into a data stream from marine industrial platform stations and a data stream from land stations, including: Based on the station code, the deployment environment of each station is identified. Data collected by stations deployed on near-shore industrial platforms is divided into the marine industrial platform station data stream, and data collected by conventional seismic observation stations deployed on bedrock or soil on land is divided into the land station data stream.

3. The method for detecting and locating differential seismic activity between land and sea for near-shore industrial platforms according to claim 1, characterized in that, The step of calling different deep learning phase picking models to perform P-wave and S-wave phase picking based on the noise domain characteristics of the data streams from the marine industrial platform stations and the land-based stations includes: Based on the data stream from the marine industrial platform stations, a first deep learning model with global context awareness and trained or transferred by real marine observation noise data is invoked to pick up seismic phases. Based on the data stream from the land stations, a second deep learning model trained on earthquake data from China and adapted for the observation area through transfer learning or parameter fitting is invoked to pick up seismic phases.

4. The method for detecting and locating differential seismic activity between land and sea for near-shore industrial platforms according to claim 3, characterized in that, The first deep learning model is a phase picking model based on the Transformer architecture and equipped with a global self-attention mechanism. It segments the data stream of the marine industrial platform station with a first sliding time window of a preset length of 60 seconds. The second deep learning model is a phase picking model based on a local feature extraction architecture and adapted to regional data. It segments the land station data stream with a second sliding time window of a preset length of 30 seconds. A preset overlap area is set between adjacent sliding time windows, and duplicate candidate phases in the overlap area are merged.

5. A method for detecting and locating differential seismic activity between land and sea for near-shore industrial platforms according to claim 1, characterized in that, The candidate seismic phase sets are aggregated into a hybrid data pool, input into a spatiotemporal seismic correlation algorithm, and combined with a regional crustal velocity model to obtain a set of associated seismic phases for valid earthquake events, including: The candidate seismic phase set is aggregated into a hybrid candidate data pool. Four-dimensional spatiotemporal seismic correlation is performed based on theoretical travel time constraints generated by the regional crustal velocity model. Simultaneously, constraints are applied on the total number of correlated seismic phases, the number of P-wave seismic phases, the number of S-wave seismic phases, the number of stations with dual seismic phases, and the travel time residual tolerance to obtain the set of correlated seismic phases of effective earthquake events. The regional crustal velocity model includes multiple velocity nodes or stratigraphic interfaces set along the depth direction, and each velocity node or stratigraphic interface is configured with P-wave velocity parameters and S-wave velocity parameters. During the correlation process, multidimensional strong constraints are applied, requiring that effective earthquake events simultaneously satisfy the following conditions: the total number of correlated seismic phases is not less than the first preset number, the number of P-wave seismic phases is not less than the second preset number, the number of S-wave seismic phases is not less than the third preset number, and at least a fourth preset number of stations simultaneously record P-waves and S-waves, and the difference between the arrival time of the correlated seismic phases and the corresponding theoretical travel time is within the preset travel time residual tolerance.

6. The method for detecting and locating differential seismic activity between land and sea for near-shore industrial platforms according to claim 1, characterized in that, Based on the arrival time data of the associated seismic phases in the aforementioned set, the nonlinear probabilistic seismic location method is used to invert the earthquake occurrence time and spatial location of the epicenter, outputting the preliminary epicenter location and location uncertainty, including: Based on the arrival time data of the associated seismic phases in the set of associated seismic phases, the posterior probability density function of the source parameters is searched in the three-dimensional source space and the range of the time of origin, so that the seismic phases of the marine industrial platform station and the seismic phases of the land station can participate in the localization together, and the maximum likelihood source location, the time of origin, and the source location uncertainty characterized by a preset confidence level are output. Based on the maximum likelihood source location, the time of origin, and the source location uncertainty, the preliminary source location and positioning uncertainty are obtained.

7. A method for detecting and locating differential seismic activity between land and sea for near-shore industrial platforms according to claim 1, characterized in that, Based on the aforementioned location uncertainty, a joint mass screening is performed on the preliminary hypocenter locations to obtain the final location results. These final location results are then compiled into a nearshore earthquake catalog, which includes: The preliminary source location is screened for quality based on at least one of the following indicators: maximum azimuth gap, root mean square travel time residual, semi-major axis of horizontal error ellipsoid, semi-minor axis of horizontal error ellipsoid, and semi-vertical error axis. The location results that meet the quality screening criteria will be output as a high-resolution catalog of nearshore earthquakes, including the time of occurrence, the latitude and longitude of the epicenter, and the focal depth.

8. A land-sea differentiated seismic detection and positioning system for near-shore industrial platforms, used to implement the method as described in any one of claims 1-7, characterized in that, include: The module includes a data acquisition and distribution module, a differentiated seismic phase picking module, a candidate seismic phase convergence module, a multidimensional strongly constrained correlation module, a nonlinear probabilistic localization module, a catalog output module, and a quality assessment and catalog output module. The data acquisition and distribution module is used to distribute the collected three-component continuous seismic waveform data into marine industrial platform station data streams and land station data streams based on the deployment environment type represented by the station metadata of each station in the marine-land hybrid network. The differentiated seismic phase picking module is used to call different deep learning seismic phase picking models to pick P-wave and S-wave seismic phases according to the noise domain characteristics of the data streams of the marine industrial platform stations and the data streams of the land stations, and to obtain a set of candidate seismic phases. The candidate seismic phases include at least the station identifier, deployment environment type, seismic phase type, seismic phase arrival time and picking probability. The candidate seismic phase aggregation module is used to aggregate the candidate seismic phase set into a hybrid data pool, input the spatiotemporal seismic correlation algorithm and combine it with the regional crustal velocity model to obtain the associated seismic phase set of effective seismic events; The nonlinear probabilistic positioning module is used to invert the earthquake occurrence time and spatial location of the source using the nonlinear probabilistic seismic positioning method based on the earthquake arrival time data in the associated seismic phase set, and output the preliminary source location and positioning uncertainty; based on the positioning uncertainty, the preliminary source location is subjected to joint quality screening to obtain the final positioning result, and the final positioning result is summarized to form a nearshore earthquake catalog.

Images