A method, system, and medium for predicting earthquake source location using TLF-GV signal correlation.

Abstract

This invention provides a method, system, and medium for predicting earthquake source locations using TLF-GV signals, belonging to the field of earthquake prediction technology. The method includes: constructing a U-Net++ model, where multiple encoding blocks are used to sequentially generate feature maps of different spatial scales based on the gradient tensors extracted from TLF-GV; multiple decoding blocks are used to decode and output earthquake source probability heatmaps based on the feature maps at different spatial scales; obtaining active fault distribution maps, and constructing a geological prior penalty term based on the distance from each grid cell in the earthquake source probability heatmap to the nearest active fault, and constructing a loss function by combining the cross-entropy loss between the probability heatmap and the actual source mask; constructing a training set, using the concatenated features of the gradient tensors extracted from the gravity gradient grid and the fault binary mask as input to train and obtain the prediction model. This method utilizes the spatial grid features of the TLF-GV gravity gradient, combined with the small-scale feature capture capability of U-Net++, to effectively improve the accuracy of earthquake source location prediction.

Description

Technical Field

[0001] This invention relates to the field of earthquake early warning technology, specifically to a method, system, and medium for predicting earthquake source locations using TLF-GV signal correlation. Background Technology

[0002] Earthquake focal location prediction is the core spatial component of short-term earthquake prediction systems. Its accuracy directly determines the targeting of disaster emergency response and the efficiency of resource allocation. Currently, the technology in this field has always revolved around traditional precursor signals (such as seismic wave arrival time, GNSS crustal deformation field, spatial distribution of underground fluids, microseismic activity sequences, etc.), and a technical system for new precursor signals with "specific frequency domain response and spatial gradient characteristics" (such as the extremely low frequency gravity-vibration signal TLF-GV) has not yet been formed, and there is a significant gap in related research. From the perspective of technological development, the early mainstream method was the time difference-of-arrival (TDA) positioning method (such as the Wadati method and the HYPODD algorithm), which inverted the source by combining the time difference of arrival of seismic waves from multiple stations with a velocity model. Although it could achieve an error of <5km in densely networked areas, the error often exceeded 15km in sparsely networked areas (interval between stations >50km). Another type of empirical formula relied on the spatial attenuation law of precursor signals (such as the negative correlation between GNSS deformation and epicentral distance) to estimate the source, but it was limited by the assumption of "stable signal attenuation". Under complex geological processes, the deviation could reach more than 20km. Both types of traditional methods could hardly meet the spatial accuracy requirements of short-term forecasts.

[0003] With the penetration of machine learning technology, earthquake source location prediction has shifted to a data-driven approach: Convolutional Neural Networks (CNNs) process two-dimensional features of GNSS deformation fields to output earthquake source probability maps, Long Short-Term Memory Networks (LSTMs) combine microseismic spatiotemporal data to optimize regional probability distributions, and ordinary U-Nets integrate underground fluid and geomagnetic data to improve resolution. Although these methods are superior to empirical methods in traditional precursor applications, their spatial features are fragmented, relying heavily on single-point data or sparse grids, failing to explore high-dimensional features such as spatial gradients and tensors, and thus failing to capture the fine coupling relationship between precursors and earthquake sources, resulting in large positioning errors. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method for predicting earthquake source locations using TLF-GV signal correlation. This method leverages the spatial grid features of the TLF-GV gravity gradient, combined with the small-scale feature capture capabilities of U-Net++, to effectively improve the accuracy of earthquake source location prediction.

[0005] To achieve the above objectives, the present invention provides the following technical solution.

[0006] A method for predicting earthquake source location associated with TLF-GV signals includes the following steps:

[0007] A U-Net++ model is constructed, comprising multiple densely connected encoding and decoding blocks. The encoding blocks are used to sequentially generate feature maps of different spatial scales based on the gradient tensors extracted from the gravity gradient grid of the low-frequency gravity-vibration signal TLF-GV. The decoding blocks are used to decode and output a probability heatmap of the earthquake source based on the feature maps of different spatial scales. The TLF-GV contains the absolute amplitude of the gravity component, the relative amplitude of the vibration component, and the signal phase.

[0008] Obtain the distribution map of active faults, convert it into a binary mask with the same resolution as the gravity gradient grid, and construct a geological prior penalty term based on the distance of each grid cell in the probability heat map to the nearest active fault. Combine the cross-entropy loss between the probability heat map and the real source mask to construct a loss function.

[0009] A gravity gradient grid, fault binary mask, and corresponding real source mask of historical earthquakes are constructed as training sets. The gradient tensor extracted from the gravity gradient grid and the concatenated features of the fault binary mask are used as input to train and obtain an earthquake source location prediction model.

[0010] Preferably, the construction of the loss function includes the following steps:

[0011] A geological prior penalty term is constructed based on the distance from each grid cell to the nearest active fault in the probability heatmap of the earthquake source.

[0012] The binary cross-entropy loss is determined based on the probability heatmap output by the model and the true source mask of the training set, and the composite loss function is determined by combining the geological prior penalty term:

[0013] ;

[0014] In the formula: For binary cross-entropy loss, This is a probabilistic heatmap output by the earthquake focal location prediction model. α is the true source mask for historical earthquakes; α is the fault constraint strength coefficient, which controls the degree of influence of geological constraints. For grid cells in a probabilistic heatmap Distance to the nearest active fault H , W Defines the grid height and width.

[0015] Preferably, the U-Net++ model includes 4 encoding blocks and 4 decoding blocks;

[0016] Each of the four coding blocks contains two 3×3 convolutional layers and one 2×2 max pooling layer, and the convolutional layers use the ReLU activation function;

[0017] Each of the four decoding blocks contains two 3×3 convolutional layers and one 2×2 transposed convolutional layer, which are connected to the corresponding scale coding blocks through dense connections, and output a probability heatmap through an activation function.

[0018] During the training of the earthquake source location prediction model, an additional channel is added; the fault binary mask and the gradient tensor extracted by the gravity gradient grid are concatenated and input into the encoding block through the additional channel.

[0019] Preferably, the extraction of the gradient tensor includes the following steps:

[0020] Acquire gradient data in the x, y, and z directions from the TLF-GV signal, and perform Z-score normalization on the gradient data in the x, y, and z directions respectively;

[0021] The gradient data after Z-score normalization are concatenated along the channel dimension to generate a 2.5D input tensor with a shape of H×W×3, where H and W are the grid height and width.

[0022] Preferably, the method further includes extracting the potential influence zone of the earthquake source based on the probability heat map output by the earthquake source location prediction model;

[0023] Extracting multi-source feature vectors and physical feature vectors from TLF-GV signals to predict the predicted magnitude. ;

[0024] Based on predicted magnitude Determine the core area of ​​the heat map: R = 20 + 5 × ( -5.0).

[0025] Preferably, the method further includes post-processing based on the probability heatmap output by the earthquake source location prediction model, specifically including:

[0026] Gravity-Vibration Contribution Index (GVCI) time series was extracted from the TLF-GV signal; the probability of earthquake occurrence (P) was predicted based on the GVCI time series. earthquake ;

[0027] Where, if P earthquake A value <0.5 is marked as a low-confidence region, and a weight is assigned to reduce the earthquake source probability value predicted by the model.

[0028] This invention also provides a TLF-GV signal-correlated earthquake source location prediction system, the system comprising:

[0029] processor;

[0030] A memory on which computer programs that can run on the processor are stored;

[0031] The computer program, when executed by the processor, implements the steps of the TLF-GV signal-correlated earthquake source location prediction method.

[0032] The present invention also provides a computer-readable storage medium storing a data processing program, which, when executed by a processor, implements the steps of the TLF-GV signal-correlated earthquake source location prediction method.

[0033] The beneficial effects of this invention are:

[0034] This invention proposes a method for predicting earthquake source locations using TLF-GV signal correlation. This method utilizes the spatial grid features of TLF-GV gravity gradients (rather than single-point data), combined with the small-scale feature capture capabilities of U-Net++, and employs a fault constraint penalty term to improve the matching rate between high-probability source areas and active faults, avoiding unreasonable results where the source is far from the fault. The location reliability conforms to geological common sense, effectively improving the accuracy of earthquake source location prediction. Attached Figure Description

[0035] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0037] Example 1

[0038] Currently, for earthquake source location prediction, the published patent "A Rapid Earthquake Response Location Method and System" (Publication No.: CN117607955A) proposes a "CNN-based source location scheme based on seismic wave signals." The core of this invention patent is to apply convolutional neural networks (CNN) to the source location of traditional seismic wave signals, attempting to optimize the location efficiency and accuracy through deep learning. Its technical logic and implementation details have clear authenticity and traceability.

[0039] The specific implementation path of the scheme is as follows: First, the P-wave, S-wave, and other seismic wave waveform sequences generated after an earthquake are collected by multiple seismic monitoring devices. Amplitude features (such as the absolute value of the maximum amplitude and the average amplitude within the sliding window) and phase features (phase correlation coefficient calculated by standard wavelet matching) are extracted from the waveform data. After denoising (wavelet threshold denoising) and standardization of the extracted features, they are input into a pre-trained CNN model. This model uses the "seismic wave features - real source coordinates" of historical earthquakes as samples and establishes a mapping relationship through end-to-end learning. In the prediction stage, the model receives the wave pattern features of the earthquake to be located and directly outputs the latitude, longitude, and depth coordinates of the source. The core advantage is that compared with the traditional time difference positioning method (which relies on manual picking of wave arrival time, taking minutes), the positioning speed is shortened to seconds, and the results can still be output when some network data is missing, which alleviates the dependence on dense network to a certain extent.

[0040] However, this scheme has fundamental flaws in terms of spatial feature utilization, geological constraint integration, technology chain synergy, and timeliness. These flaws are precisely the core improvement directions of this invention, and their specific manifestations and causes are as follows:

[0041] First, the spatial feature dimension is lacking, failing to capture the fine correlation of the source area. This scheme relies only on the single-point temporal features of seismic waves (such as the amplitude time series and phase changes of a single station), without constructing spatial grid features of multiple stations (such as wave velocity differences and amplitude gradient distribution between different stations). The CNN model input is one-dimensional temporal data, which cannot learn the coupling relationship between "spatial features and source location," resulting in an output of only a single coordinate point, failing to reflect the location uncertainty, and with an error exceeding 15km when the inter-station spacing is >30km. Second, the lack of geological structural constraints leads to location results that defy common sense. The scheme uses pure data-driven modeling, and the loss function only optimizes the "Euclidean distance between the predicted coordinates and the actual source," without incorporating the rule that "earthquakes mostly occur in active fault zones." It neither uses fault distribution as a spatial prior for training data nor penalizes predictions far from faults, resulting in an unreasonable location rate (such as locating within stable blocks) of up to 23% in practical applications. Third, the technology chain is fragmented, wasting data value. The input to the scheme is limited to seismic wave characteristics and does not incorporate the results of previous stages such as "magnitude prediction" and "signal reliability assessment." For example, it does not utilize magnitude parameters to adjust the spatial range of the location (the larger the magnitude, the larger the radius of influence of the epicenter should be), resulting in the location results of small-magnitude earthquakes including a large number of irrelevant areas. Fourth, the timeliness is extremely low and cannot meet the needs of short-term forecasting. This is the most critical weakness of the scheme, stemming from its reliance on the essential attribute of "seismic wave signals": seismic waves (P-waves and S-waves) only generate and propagate after an earthquake, resulting in a natural propagation delay. For example, at a distance of 100km from the epicenter and a focal depth of 10km, P-waves (velocities of 6-8km / s) take 12.5-16.7 seconds to reach the station, while S-waves (velocities of 3-4km / s) take 25-33.3 seconds. Furthermore, the scheme requires at least 1-2 seconds of waveform data to extract effective features (avoiding noise interference from individual points), and then preprocessing (0.5-1 seconds) and model inference (0.1-0.5 seconds). Overall, the time from earthquake occurrence to outputting the location result is 15-40 seconds. By this time, the S-wave has already reached the vicinity of the epicenter, and the disaster has begun. The scheme can only achieve "rapid post-earthquake location," not "short-term pre-earthquake prediction," completely failing to provide time for emergency response (short-term forecasts require several minutes to several hours in advance). To address this, this invention proposes a method for predicting earthquake source location based on TLF-GV signal correlation. This method leverages the characteristics of TLF-GV signals (1-30 mHz), which exhibit anomalies (such as changes in gravity gradient tensor and GVCI exponential fluctuations) several hours to days before an earthquake. This allows for the early detection of precursors without waiting for the arrival of seismic waves. Combined with a fault-constrained U-Net++ model, the process from anomaly signal processing to outputting a probability heatmap takes only a few minutes, providing a sufficient response window for short-term forecasts and completely overcoming the timeliness deficiencies of existing seismic wave location schemes.

[0042] like Figure 1 As shown, the main steps include:

[0043] S1: Construct a U-Net++ model, including multiple densely connected encoding blocks and decoding blocks; the multiple encoding blocks are used to generate feature maps of different spatial scales in sequence based on the gradient tensors extracted from the gravity gradient grid of the low-frequency gravity-vibration signal TLF-GV; the multiple decoding blocks are used to decode and output the probability heat map of the earthquake source based on the feature maps of different spatial scales.

[0044] S2: Obtain the distribution map of active faults, convert it into a binary mask with the same resolution as the gravity gradient grid, and construct a geological prior penalty term based on the distance of each grid cell in the probability heat map to the nearest active fault. Combine the cross-entropy loss between the probability heat map and the real source mask to construct a loss function.

[0045] S3: Construct a gravity gradient grid, fault binary mask, and corresponding real source mask of historical earthquakes as a training set. Use the gradient tensor extracted from the gravity gradient grid and the features of the fault binary mask as input to train and obtain an earthquake source location prediction model.

[0046] The specific steps are as follows:

[0047] 1. Step 1: Input Data Preprocessing

[0048] 1.1 Data Reception: Receiving three types of core data:

[0049] ①TLF-GV gravity gradient tensor grid (from signal processing stage):

[0050] It contains gradient data in the x-axis (∂g / ∂x), y-axis (∂g / ∂y), and z-axis (∂g / ∂z), with a grid resolution of 5km×5km and a coverage area of ​​≥500km×500km. The data format is GeoTIFF (supporting spatial coordinate association).

[0051] The TLF-GV dataset of low-frequency gravity-vibration signals contains data such as the absolute amplitude of gravity components, the relative amplitude of vibration components, and signal phase. It has three characteristics: ultra-low frequency (1~30mHz), micro-amplitude (1~10μGal), and short-term (occurring 1~20 days before an earthquake). It is collected by deploying atmospheric tidal gravimeters (main layer, relative gravity, wide-area capture) and solid tidal gravimeters (secondary layer, relative gravity, gravity component supplement) in accordance with risk levels, combined with a dual-mode collaborative cold atom interferometer. The specific acquisition method is disclosed by the team of this invention in a patent application filed on the same day, a method for network monitoring of short-term earthquake precursor information.

[0052] ② Active fault distribution map (from the National Earthquake Science Database): vector format (SHP file), with fault strike, dip angle, and activity level marked (e.g., "Holocene active faults" are high priority);

[0053] ③ Preliminary prediction auxiliary information (from the time / magnitude prediction stage): Probability of earthquake occurrence (P) earthquake (e.g., 0.85), predicted magnitude ( (e.g., 6.3), used to constrain the range and reliability of heatmaps.

[0054] Among them, the method of extracting multi-source feature vectors and physical feature vectors from TLF-GV signals to predict the predicted magnitude is disclosed in a patent application filed on the same day by the invention team: a method, system, and medium for predicting earthquake magnitude using TLF-GV signals. Specifically, the gravity-vibration contribution index (GVCI) time series is extracted from the TLF-GV signal, and the probability of earthquake occurrence (P) is predicted using the GVCI time series. earthquake The aforementioned technical means were disclosed by the invention team in a patent application filed on the same day: a method and system for predicting earthquake occurrence time.

[0055] 1.2 Data Preprocessing:

[0056] ① Gravity gradient grid normalization:

[0057] Z-score standardization (Gnorm=(G-μG) / σG, where μG and σG are the mean and standard deviation of the TLF-GV gradient during the quiescent period) is performed on the gradients in the x, y, and z directions to eliminate dimensional differences.

[0058] Construct a 2.5D input tensor: Concatenate the normalized gradients in the x / y / z directions along the channel dimension (shape is H×W×3, where H and W are the grid height / width, such as 100×100 pixels). Each pixel corresponds to a 5km×5km spatial unit, and the pixel value reflects the TLF-GV gradient anomaly intensity of that unit.

[0059] ② Vectorization of fault data:

[0060] The SHP format tomographic map is converted into a binary mask (H×W×1) with the same resolution as the gradient grid: the pixel values ​​within 5km of the fault line are set to 1 ("constructing the constraint area"), and the rest are set to 0 ("unconstrained area"), which is used for fault constraint injection in subsequent models.

[0061] ③ Auxiliary information adaptation:

[0062] Based on predicted magnitude Determine the core area of ​​the heat map: R = 20 + 5 × ( -5.0) (unit: km), such as ( When =6.3), R=26.5km, ensuring the heat map focuses on the potential impact zone of the earthquake source;

[0063] If Pearthquake < 0.5 (low probability of earthquake), mark the "low confidence region" (such as the edge grid) in the gradient grid, and reduce the weight of the region in subsequent predictions.

[0064] 1.3 Data Validation: Ensure that the preprocessed data meets the following requirements: no missing values ​​in the gradient grid (missing rate ≤2%, filled by Kriging interpolation), and the fault mask is aligned with the grid spatial coordinates (error ≤1 pixel) to avoid constraint failure caused by spatial misalignment.

[0065] 2. Step 2: Fault Constraint U-Net++ Model Construction

[0066] 2.1 U-Net++ Core Network Design (Adapted to TLF-GV Gradient Mesh)

[0067] U-Net++ is an improved version of U-Net. It enhances the capture capability of small-scale features (such as TLF-GV gradient anomaly clustering regions) through a densely connected encoder-decoder structure, and is adapted to spatial grids with a resolution of 5km×5km.

[0068] 2.1.1 Encoder (Feature Extraction):

[0069] ① Structure: 4 encoding blocks, each block contains 2 3×3 convolutional layers (with ReLU activation function) and 1 2×2 max pooling layer (with stride of 2 and downsampling);

[0070] ② Function: Extract TLF-GV spatial features from 2.5D gradient tensor - shallow coding blocks (layers 1-2) capture local gradient anomalies (such as single-point high gradients), and deep coding blocks (layers 3-4) capture global features (such as spatial clustering of gradient anomalies, corresponding to stress concentration in the seismic source area).

[0071] ③ Output: Feature maps at 4 scales (e.g., 100×100→50×50→25×25→12×12), retaining TLF-GV spatial information at different scales.

[0072] 2.1.2 Decoder (Feature Recovery and Heatmap Generation):

[0073] ① Structure: 4 decoding blocks, each block contains 2 3×3 convolutional layers and 1 2×2 transposed convolutional layer (stride 2, upsampling), which are fused with encoder feature maps of the corresponding scale through dense connections;

[0074] ② Function: Gradually restore the TLF-GV features extracted by the encoder to the original grid resolution and generate an H×W×1 probability heatmap (pixel value 0~1, representing the probability that the 5km×5km unit is the source).

[0075] ③ Output activation function: Sigmoid (ensures the probability value is between 0 and 1, which facilitates spatial probability interpretation).

[0076] 2.2 Fault-constrained injection

[0077] A geological prior penalty term is added to the model loss function to force the predicted high-probability seismic source areas to align with active faults, thus avoiding violations of tectonic laws. The composite loss function is defined as follows:

[0078]

[0079] In the formula: For binary cross-entropy loss, This is the probability heatmap output by the model. The "true source mask" for historical earthquakes (pixel value of 1 within 10km of the true source, and 0 elsewhere) ensures the accuracy of data-driven positioning; α is the fault constraint strength coefficient (optimized to α=0.3 through validation set) to control the degree of influence of geological constraints. For grid cells The distance to the nearest active fault (unit: pixel, 1 pixel = 5km) increases with distance; the greater the distance, the larger the penalty. (The probability of a seismic source in this unit is) high, but it is far from the fault. If the value is large, the penalty term increases significantly, forcing the model to reduce the probability of that unit.

[0080] Physical significance: By penalizing predictions that are "high probability but far from faults", the high probability zone of the earthquake source is ensured to be consistent with the active fault zone, which is in line with the geological law that "earthquakes mostly occur in fault zones".

[0081] In addition, fault feature fusion is added during training: the fault binary mask generated in step 1.2 is used as an additional channel (H×W×1) and concatenated with the TLF-GV gradient tensor (H×W×3) to form the input of H×W×4, so that the model can directly learn the correlation between "TLF-GV gradient anomaly + fault location".

[0082] 2.3 Model Training and Optimization

[0083] 2.3.1 Construction of training dataset:

[0084] ① Training set: 100+ historical earthquake samples with "TLF-GV gradient grid + fault mask + real source mask" (e.g., the 7.8 magnitude earthquake in Turkey: gradient grid covers 1000km×1000km, and real source mask marks the area 10km around the epicenter).

[0085] ② Validation set / Test set: 20% and 10% of the historical samples are selected respectively. The validation set is used to fine-tune hyperparameters (such as α, network depth), and the test set is used to evaluate the positioning accuracy (such as "the proportion of high probability areas containing real earthquake sources" and "source center error").

[0086] 2.3.2 Hyperparameter Tuning: The optimal parameters were determined using grid search as follows:

[0087] Table 1. Optimal values ​​of hyperparameters;

[0088]

[0089] 2.3.3 Model Evaluation Criteria: The test set must meet the following requirements:

[0090] ① The proportion of real earthquake sources in high-probability areas (probability ≥ 0.6) is ≥ 90%;

[0091] ② The source center error (the distance between the predicted high-probability cluster center and the actual epicenter) is ≤5km; if it is not met, readjust (e.g., increase α or increase the number of training rounds).

[0092] 3. Step 3: Source Probability Prediction and Post-processing

[0093] 3.1 Source Probability Prediction: Input the preprocessed "TLF-GV gradient grid + fault mask" into the trained model, and output a source probability heatmap of H×W×1. Example: A grid cell (N30.6°, E103.3°) has a probability value of 0.75, which means that the probability of this 5km×5km cell being the source is 75%, and the probability of surrounding cells decreases with distance.

[0094] 3.2 Post-processing of heatmaps:

[0095] ① Scope constraint: Based on the magnitude predicted in the preceding earthquake. Extract the core area of ​​the heatmap (R=20+5×( -5.0), eliminating low-probability edge units (probability <0.2) and focusing on potential seismic source areas;

[0096] ② Spatial clustering to extract high-probability areas: The DBSCAN clustering algorithm is used (neighborhood radius = 2 pixels (10km), minimum number of points = 3). The neighborhood radius adopts the magnitude adaptive formula: R = 5 + M × 1 (unit: km), where M is the preceding predicted magnitude; for example, when M = 6.3, R = 11.3km (approximately 2.26 pixels). The minimum number of points is simultaneously adjusted to 4 to ensure that the clustering range matches the magnitude influence radius, and the high-probability area covers ≥90% of the actual earthquake source. Continuous units with a probability ≥ 0.6 are clustered into "high-probability source areas," and the center coordinates (e.g., "N30.58°, E103.25°") and area (e.g., 25km × 25km) of each cluster are output.

[0097] ③ Credibility Correction: If the probability of the preceding earthquake is P earthquake If the probability is less than 0.5, multiply all unit probabilities by 0.8 (to reduce overall confidence) and mark it as "low confidence location, TLF-GV signal needs to be continuously monitored".

[0098] 3.3 Result verification: Compare the positional relationship between the high-probability source area and the active fault. If the high-probability area is more than 10km away from the nearest fault (violating tectonic constraints), then re-examine the TLF-GV gradient data (e.g., whether there are measurement errors) and re-predict if necessary.

[0099] 4. Step 4: Output of Structured Report

[0100] (1) Objective: To generate reports containing spatial visualization and quantitative indicators to support joint forecasting and early warning decision-making based on the three elements.

[0101] (2) Report content (in JSON+PDF format, including heatmap visualization):

[0102] ① Enter basic information:

[0103] TLF-GV gradient grid range (e.g., "N29.0°~N32.0°, E101.0°~E104.0°"), resolution (5km×5km);

[0104] Preceding auxiliary information (probability of earthquake occurrence 0.85, predicted magnitude 6.3).

[0105] ②Location results:

[0106] Heatmap of earthquake focal probability (GeoTIFF format, can be overlaid on a map);

[0107] List of high-probability earthquake source areas (e.g., "Region 1: N30.58°~N30.68°, E103.20°~E103.30°, probability 0.6~0.75, area 25km×25km").

[0108] The coordinates of the epicenter (e.g., "Center of the main high probability zone: N30.63°, E103.25°").

[0109] ③Constraint satisfaction:

[0110] The distance between the high-probability zone and the nearest fault (e.g., "2.5km, in line with tectonic constraints").

[0111] The effect of preceding magnitude on the range (e.g., "core range 26.5 km, matching magnitude 6.3").

[0112] ④ Risk advice:

[0113] For high-probability areas (probability ≥ 0.6), it is recommended to activate key early warning systems;

[0114] In the medium probability region (0.4 ≤ probability < 0.6), it is recommended to strengthen the monitoring of TLF-GV signals.

[0115] The present invention has the following advantages over the prior art:

[0116] 1. Make full use of TLF-GV spatial characteristics: mine the source area information contained in the gravity gradient tensor grid data (2.5D spatial distribution) of TLF-GV signal, upgrade the positioning data dimension from the traditional "single point time series" to "multi-site spatial grid", and reduce the positioning error from more than 10km to less than 5km.

[0117] 2. Injecting geological fault constraints: Active fault distribution data are incorporated into model training and prediction. Predictions far from faults are penalized by a loss function to ensure that high-probability seismic source areas are consistent with tectonically active areas, reducing the unreasonable location rate to below 5%.

[0118] 3. Output spatial probability heat map: In response to the risk zoning requirements of short-term forecasts, output a 2.5D probability heat map of the source area (such as "probability value of a 5km×5km grid") instead of a single coordinate point, to support the hierarchical decision-making of "key early warning for high probability areas and continuous monitoring for low probability areas".

[0119] 4. Adapt to preceding technology chain: Directly accept gravity gradient data, earthquake time probability information, and magnitude range parameters from TLF-GV signal processing without additional data conversion, ensuring technical synergy between "TLF-GV networking - signal processing - time / magnitude prediction - earthquake source location".

[0120] The above is an embodiment of the TLF-GV signal-correlated earthquake source location prediction method provided in this example. Based on the same idea, this embodiment also provides a corresponding TLF-GV signal-correlated earthquake source location prediction system. Specific limitations of the TLF-GV signal-correlated earthquake source location prediction system can be found in the limitations of the TLF-GV signal-correlated earthquake source location prediction method described above, and will not be repeated here. Each module in the above TLF-GV signal-correlated earthquake source location prediction system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0121] This embodiment also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1The provided method for predicting earthquake source locations by associating TLF-GV signals.

[0122] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0123] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting earthquake source location using TLF-GV signal correlation, characterized in that, Includes the following steps: A U-Net++ model is constructed, comprising multiple densely connected encoding and decoding blocks; the multiple encoding blocks are used to sequentially generate feature maps of different spatial scales based on the gradient tensors extracted from the gravity gradient grid of the low-frequency gravity-vibration signal TLF-GV. Multiple decoding blocks are used to decode and output a probabilistic heatmap of the earthquake source based on feature maps at different spatial scales; wherein, the TLF-GV contains the absolute amplitude of the gravity component, the relative amplitude of the vibration component, and the signal phase; Obtain the distribution map of active faults, convert it into a binary mask with the same resolution as the gravity gradient grid, and construct a geological prior penalty term based on the distance of each grid cell in the probability heat map to the nearest active fault. Combine the cross-entropy loss between the probability heat map and the real source mask to construct a loss function. A gravity gradient grid, fault binary mask, and corresponding real source mask of historical earthquakes are constructed as training sets. The gradient tensor extracted from the gravity gradient grid and the concatenated features of the fault binary mask are used as input to train and obtain an earthquake source location prediction model. The construction of the loss function includes the following steps: A geological prior penalty term is constructed based on the distance from each grid cell to the nearest active fault in the probability heatmap of the earthquake source. The binary cross-entropy loss is determined based on the probability heatmap output by the model and the true source mask of the training set, and the composite loss function is determined by combining the geological prior penalty term: ； In the formula: For binary cross-entropy loss, This is a probabilistic heatmap output by the earthquake focal location prediction model. α is the true source mask for historical earthquakes; α is the fault constraint strength coefficient, which controls the degree of influence of geological constraints. For grid cells in a probabilistic heatmap Distance to the nearest active fault H , W Defines the grid height and width.

2. The method for predicting earthquake source location by association of TLF-GV signals according to claim 1, characterized in that, The U-Net++ model includes 4 encoding blocks and 4 decoding blocks; Each of the four coding blocks contains two 3×3 convolutional layers and one 2×2 max pooling layer, and the convolutional layers use the ReLU activation function; Each of the four decoding blocks contains two 3×3 convolutional layers and one 2×2 transposed convolutional layer, which are connected to the corresponding scale coding blocks through dense connections, and output a probability heatmap through an activation function. During the training of the earthquake source location prediction model, an additional channel is added; the fault binary mask and the gradient tensor extracted by the gravity gradient grid are concatenated and input into the encoding block through the additional channel.

3. The method for predicting earthquake source location by association of TLF-GV signals according to claim 1, characterized in that, The extraction of the gradient tensor includes the following steps: Acquire gradient data in the x, y, and z directions from the TLF-GV signal, and perform Z-score normalization on the gradient data in the x, y, and z directions respectively; The gradient data after Z-score normalization are concatenated along the channel dimension to generate a 2.5D input tensor.

4. The method for predicting earthquake source location by association of TLF-GV signals according to claim 1, characterized in that, It also includes extracting the potential influence zone of the earthquake source based on the probability heat map output by the earthquake source location prediction model; Extracting multi-source feature vectors and physical feature vectors from TLF-GV signals to predict the predicted magnitude. ; Based on predicted magnitude Determine the core area of ​​the heat map: R = 20 + 5 × ( -5.0).

5. The method for predicting earthquake source location by association of TLF-GV signals according to claim 1, characterized in that, It also includes post-processing of the probability heatmap output by the earthquake focal location prediction model, specifically including: Gravity-Vibration Contribution Index (GVCI) time series was extracted from the TLF-GV signal; the probability of earthquake occurrence (P) was predicted based on the GVCI time series. earthquake ; Where, if P earthquake A value <0.5 is marked as a low-confidence region, and a weight is assigned to reduce the earthquake source probability value predicted by the model.

6. A TLF-GV signal-correlated earthquake source location prediction system, characterized in that, The system includes: processor; A memory on which computer programs that can run on the processor are stored; When the computer program is executed by the processor, it implements the steps of the TLF-GV signal-associated earthquake source location prediction method as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a data processing program, which, when executed by a processor, implements the steps of the TLF-GV signal-associated earthquake source location prediction method as described in any one of claims 1 to 5.

Images