Single-station seismic positioning system and method combining multi-source features and collaborative loss

Abstract

The invention provides a single-station earthquake positioning system and method combining multi-source features and collaborative loss, and relates to the technical field of earthquake positioning, and the single-station earthquake positioning system comprises a data preprocessing module, a model training module, and a single seismic source positioning model. Inputting three-component waveform data into a data preprocessing module by querying seismic directory information, inputting the preprocessed seismic waveform data into a single seismic source positioning model, training each network module to obtain a single-station seismic positioning system, and outputting a time sequence feature tensor, a time frequency feature tensor and a station information feature tensor; the trained seismic positioning system is used for predicting the seismic source depth and the epicentral distance, seismic waveform data collected by the station are input into the data preprocessing module to obtain standardized waveform data, the standardized waveform data are input into the seismic source positioning module, and prediction results of the seismic source depth and the epicentral distance of the station are output.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake positioning, and in particular to a single-station earthquake positioning system and method combining multi-source characteristics and coordinated losses. Background Art

[0002] Earthquake is a natural phenomenon in which the energy in the earth's crust is rapidly released, causing the surface to vibrate. The destructive force it produces can not only pose a serious threat to human life and property, but may also have a profound impact on geological structures and the ecological environment. Earthquake location is one of the important topics in seismological research. The main task is to determine the source location, source depth, and corresponding epicentral distance of the earthquake event. The source depth determines the depth of earthquake energy release and the degree of impact on the surface. The epicentral distance indicates the horizontal distance from the observation station to the epicenter, which is directly related to the propagation time and amplitude of the seismic wave. Accurate source location is of great significance for earthquake early warning, post-earthquake rescue, tectonic activity analysis, and earthquake disaster assessment. Current earthquake location methods usually rely on observation data from multiple seismic stations, and calculate the source location by analyzing the time difference between seismic waves arriving at different stations and the change in seismic wave amplitude. However, the distribution of seismic stations in many areas is relatively sparse, or due to terrain and environmental restrictions, earthquake location can only be based on the observation data of a single station in some special cases.

[0003] Traditional multi-station positioning relies on triangulation or inversion technology, which can accurately determine the earthquake source through multiple station observation points in space. In the case of a single station, due to the lack of multiple references in space, it is difficult to restore the spatial position and depth information of the earthquake source based on the waveform data of only one observation point. Therefore, the accuracy and stability of single-station earthquake source positioning will be greatly reduced, and it is easily limited by the limited information dimension, resulting in unsatisfactory positioning results. Most of the current single-station methods can only provide the approximate range of the earthquake source, which is difficult to meet the needs of high-precision positioning.

[0004] In recent years, the development of artificial intelligence technologies such as deep learning has brought new solutions to single-station earthquake location. Deep learning models can learn complex time series characteristics, spectral characteristics, and potential earthquake laws from a large amount of seismic observation data, which makes it possible to use single-station waveforms for earthquake source location. Summary of the invention

[0005] In view of the shortcomings of existing single-station positioning technology, the present invention provides a single-station earthquake positioning system and method that combines multi-source features and collaborative losses. Through a multi-branch deep neural network structure and a multi-station indirect collaborative training strategy, features are extracted from waveform timing features, time-frequency features, and station location information respectively. In the training stage, a custom loss function is used to simulate the collaborative effect of multiple stations, thereby improving the source prediction capability of single-station observation data.

[0006] A single-station earthquake location system combining multi-source characteristics and coordinated losses, comprising: a data preprocessing module, a model training module, and a single-station earthquake source location model;

[0007] The data preprocessing module is used to preprocess the collected original seismic waveform data to obtain the preprocessed seismic waveform data, that is, the seismic three-channel waveform data;

[0008] The model training module inputs the three-channel earthquake waveform data into the single-station source location model, trains the single-station source location model through collaborative loss, and dynamically adjusts the learning rate during the training process to obtain the trained single-station source location model;

[0009] The synergy loss is specifically as follows: on the basis of the mean square error, the synergy loss of multiple stations is added. First, the three-channel earthquake waveform data is adjusted so that each training batch contains multiple stations of the same earthquake event. Then, when calculating the loss, the epicenter distance of the station predicted by the model and the geographical location information of the station are used to estimate the location of a source through a geometric method. Finally, the spherical distance between the estimated source and the actual source is calculated as the synergy loss auxiliary model training.

[0010] The single-station earthquake source location model is a three-branch deep learning network, which inputs the earthquake waveform data collected by the station into a data preprocessing module, obtains the preprocessed earthquake waveform data, inputs it into the trained single-station earthquake source location model, and outputs the prediction results of the earthquake source depth and epicenter distance; specifically, it includes: a time series feature extraction network based on multi-scale convolution and Bi-Mamba, a time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and a station feature extraction network based on geographic information embedding;

[0011] The time series feature extraction network based on multi-scale convolution and Bi-Mamba uses a multi-scale convolution kernel to extract multi-scale features from the input three-channel seismic waveform data, and enhances the integrity of the features through residual connections; then the Bi-Mamba module is used to capture the long-term dependency of the waveform data, obtain the time series feature tensor, and realize preliminary time series modeling;

[0012] The time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention first generates a time-frequency graph through short-time Fourier transform STFT of the input three-channel seismic waveform data, and then divides the time-frequency graph into blocks using the Patch Embedding layer and inputs it into the ConvMixer network, gradually extracts multi-scale spatial and frequency features, and finally generates a time-frequency feature tensor; the time-frequency feature tensor passes through the MCANet module, and weights the key features at different scales based on multi-scale cross-axis attention to obtain a time-frequency feature tensor;

[0013] The station feature extraction network based on geographic information embedding embeds the geographic location information of the station, converts it into a feature tensor through a fully connected layer, and obtains a station information feature tensor;

[0014] The time series feature tensors, time-frequency feature tensors and station information feature tensors output by the above three networks are spliced ​​to generate a comprehensive feature tensor; the comprehensive feature tensor passes through the regression head of multiple layers of fully connected layers, and the final predicted values ​​of focal depth and epicenter distance are output with the Softplus activation function;

[0015] A single-station earthquake location system and method combining multi-source characteristics and coordinated loss includes the following steps:

[0016] Step 1: Query the earthquake catalog information to determine the arrival time of the three-component earthquake P wave; after determining the first arrival time of the P wave, intercept the data from a second before to b seconds after the starting point of each component;

[0017] Step 2: Input the obtained three-component waveform data into the data preprocessing module, remove the seismic waveform data containing discontinuities and the signal-to-noise ratio below the set threshold, divide the seismic waveform data by the sensitivity of the seismometer, and then filter, mean detrend and normalize the waveform;

[0018] Step 3: Input the preprocessed seismic waveform data into the time series feature extraction network based on multi-scale convolution and Bi-Mamba, the time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and the station feature extraction network based on geographic information embedding, train each network module, obtain a single-station earthquake positioning system, and output the time series feature tensor, time-frequency feature tensor, and station information feature tensor;

[0019] Step 3.1: Input the preprocessed seismic waveform data into the time series feature extraction network based on multi-scale convolution and Bi-Mamba, capture the features of different scales in the time series dimension through the multi-scale convolution kernel, and combine the residual connection to retain the original waveform information; then use the bidirectional Bi-Mamba module to extract the long time series dependency features and output the time series feature tensor;

[0020] Step 3.2: Input the preprocessed seismic waveform data into the time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and output the time-frequency feature tensor;

[0021] Step 3.2.1: Use short-time Fourier transform (STFT) to convert the input waveform data into a two-dimensional time-frequency diagram to associate the time domain and frequency domain information;

[0022] Step 3.2.2: The time-frequency graph is divided into blocks through the Patch Embedding layer and input into the ConvMixer structure, and the time-frequency features are gradually extracted through Depthwise convolution and Pointwise convolution;

[0023] Step 3.2.3: Input the extracted time domain features into the multi-scale cross-axis attention module MCANet, perform multi-axis weighting on the time-frequency features, and generate a time-frequency feature tensor;

[0024] Step 3.3: Input the geographical location information of the station into the station feature extraction network based on geographical information embedding, and use the fully connected layer to map the geographical information into the station information feature tensor; the geographical location information includes longitude, latitude and elevation;

[0025] Step 4: Concatenate the time series feature tensor, time-frequency feature tensor and station information feature tensor obtained in step 3 to generate a comprehensive feature tensor;

[0026] Step 5: Input the comprehensive feature tensor into the regression head, and pass through multiple layers of fully connected layers and Softplus activation function in sequence. The trained regression head outputs the predicted values ​​of focal depth and epicenter distance;

[0027]

[0028] In the formula represents the final regression prediction value of the model, Z represents the comprehensive feature tensor input to the regression head, b represents the bias parameter of the regression head, and W reg is the weight parameter of the regression head, and Softplus is the activation function to ensure that the output value is non-negative;

[0029] Step 6: Conduct model training for the single-station source location model;

[0030] Step 6.1: Add multi-station collaborative loss. Each training batch contains 64 observation data. Every four observations are input from four different stations of the same earthquake event. For the four stations of the same event, the epicenter distance predicted by the single-station source location model and the station's geographical location are used to infer the source location, and compared with the actual source location, the position deviation is calculated as an additional loss and added to the total loss function for training.

[0031] Specifically, the geographical distance between the station and the predicted earthquake source is calculated by the Haversine formula:

[0032]

[0033] Among them, R is the radius of the earth, lat1 and lat2 represent the latitude of the station and the predicted source, respectively, Δlat = lat2-lat1 represents the latitude difference between the station and the source location, Δlon = lon2-lon1 represents the longitude difference between the station and the source location; lat2 is the latitude of the source, lat1 is the latitude of the station; lon2 is the longitude of the source, lon1 is the longitude of the station, and the source location is estimated by minimizing the difference between the predicted epicenter distance of the four stations and the actual location of the station. The estimated location is compared with the actual source location, and the position deviation is calculated and added as a collaborative loss to the total loss function for optimization. The final total loss function is:

[0034] Total Loss=DepthLoss+Distance Loss+Geographical Loss

[0035] Among them, Depth Loss and Distance Loss are the mean square errors of the predicted focal depth and epicenter distance, and Geographical Loss is the coordination loss of multiple stations;

[0036] Step 6.2: Dynamically adjust the learning rate and observe the test effect during model training. When the prediction error is greater than the set threshold and the loss no longer decreases, automatically reduce the learning rate. When the model further converges and the loss begins to decrease, maintain or increase the learning rate by 0.001.

[0037] Step 7: Use the trained earthquake location system to predict the focal depth and epicenter distance. Input the earthquake waveform data collected by the station into the data preprocessing module to obtain standardized waveform data, which is input into the earthquake source location module to output the predicted results of the focal depth and epicenter distance of the station.

[0038] The beneficial effects of adopting the above technical solution are:

[0039] The present invention provides a single-station earthquake positioning system and method that combines multi-source features with collaborative losses. The present invention achieves high-precision prediction of focal depth and epicenter distance under single-station conditions through effective multi-feature fusion and collaborative optimization strategies, effectively improving the accuracy and reliability of single-station earthquake positioning. This method not only overcomes the observation challenges under sparse or environmental constraints of multiple stations, but also provides technical support for rapid response to earthquake disasters and accurate judgment of the scope of earthquake impact. The application of this technology is helpful for earthquake early warning, minimizing the loss of life and property caused by earthquakes, has important social and economic benefits, and has far-reaching practical significance for my country's earthquake prevention and disaster reduction cause. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1This is a flow chart of a single-station earthquake location method combining multi-source characteristics and coordinated loss according to an embodiment of the present invention;

[0041] Figure 2 This is a structural diagram of a single-station earthquake positioning system combining multi-source characteristics and coordinated loss according to an embodiment of the present invention; DETAILED DESCRIPTION

[0042] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0043] The software environment of this implementation is WINDOWS10 system, Anaconda+pytorch+NVIDIA1060Ti+Pycharm IDE;

[0044] A single-station earthquake location system combining multi-source characteristics and cooperative losses, such as Figure 2 As shown, it includes: data preprocessing module, model training module, single earthquake source positioning model;

[0045] The data preprocessing module is used to preprocess the collected original seismic waveform data to obtain the preprocessed seismic waveform data, that is, the seismic three-channel waveform data;

[0046] The raw data may be interfered by external factors during the acquisition process, resulting in a decrease in data quality. Waveform data containing discontinuities and low signal-to-noise ratios need to be filtered to ensure the stability and accuracy of the input data.

[0047] The model training module inputs the three-channel earthquake waveform data into the single-station earthquake source location model, trains the single-station earthquake source location model through collaborative loss, and dynamically adjusts the learning rate during the training process to improve the convergence and stability of the model; and obtains the trained single-station earthquake source location model;

[0048] The collaborative loss is specifically: adding multi-station collaborative loss on the basis of mean square error, the purpose is to use the information of multiple stations to constrain the rationality of the model's prediction of the epicentral distance of a single station during the model training process. The specific method is to first adjust the three-channel earthquake waveform data so that each training batch contains multiple stations of the same earthquake event, and then when calculating the loss, use the model-predicted station epicentral distance and the station's geographical location information to estimate the location of a source through a geometric method, and finally calculate the spherical distance between the estimated source and the actual source as the collaborative loss auxiliary model training;

[0049] The single-station earthquake source location model is a three-branch deep learning network, which inputs the earthquake waveform data collected by the station into a data preprocessing module, obtains the preprocessed earthquake waveform data, inputs it into the trained single-station earthquake source location model, and outputs the prediction results of the earthquake source depth and epicenter distance; specifically, it includes: a time series feature extraction network based on multi-scale convolution and Bi-Mamba, a time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and a station feature extraction network based on geographic information embedding;

[0050] The time series feature extraction network based on multi-scale convolution and Bi-Mamba uses a multi-scale convolution kernel to extract multi-scale features from the input three-channel seismic waveform data, and enhances the integrity of the features through residual connections; then the Bi-Mamba module is used to capture the long-term dependency of the waveform data, obtain the time series feature tensor, and realize preliminary time series modeling;

[0051] The time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention first generates a time-frequency graph through short-time Fourier transform STFT of the input three-channel seismic waveform data, and links the time domain and frequency domain information to reflect the two-dimensional characteristics of the signal; then the time-frequency graph is divided into blocks using the Patch Embedding layer and input into the ConvMixer network, and multi-scale spatial and frequency features are gradually extracted to finally generate a time-frequency feature tensor; the time-frequency feature tensor passes through the MCANet module, and the key features at different scales are weighted based on multi-scale cross-axis attention to obtain the time-frequency feature tensor, thereby enhancing the model's attention to the spectrum information;

[0052] In this embodiment, the key features are the features that are more strongly correlated with the model prediction results. Specifically, which features are black boxes to us are found by the network itself.

[0053] The station feature extraction network based on geographic information embedding embeds the geographic location information (such as longitude, latitude and elevation) of the station, converts it into a feature tensor through a fully connected layer, and obtains a station information feature tensor;

[0054] The time series feature tensors, time-frequency feature tensors and station information feature tensors output by the above three networks are spliced ​​to generate a comprehensive feature tensor; the comprehensive feature tensor passes through the regression head of multiple layers of fully connected layers, and the final predicted values ​​of focal depth and epicenter distance are output with the Softplus activation function;

[0055] The specific Softplus activation function is:

[0056] Softplus(x)=log(1+e x )

[0057] Single-station earthquake location system and method combining multi-source characteristics and coordinated loss, such as Figure 1 As shown, the following steps are included:

[0058] Step 1: Query the earthquake catalog information to determine the arrival time of the three-component earthquake P wave; after determining the first arrival time of the P wave, intercept the data from a second before to b seconds after the starting point of each component;

[0059] In this embodiment, the data from 3 seconds before to 37 seconds after the starting point of each component is intercepted, and the sampling frequency of the data is 100 Hz, with a total of 4000 sampling points;

[0060] Step 2: Input the obtained three-component waveform data into the data preprocessing module, remove the seismic waveform data containing discontinuities and the signal-to-noise ratio below the set threshold, divide the seismic waveform data by the sensitivity of the seismometer to correct the error caused by the instrument's own influencing factors, and then filter, mean detrend and normalize the waveform; in this embodiment, 3*4000 seismic waveform data is obtained, that is, three channels, 4000 sampling points per channel;

[0061] Step 3: Input the preprocessed seismic waveform data into the time series feature extraction network based on multi-scale convolution and Bi-Mamba, the time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and the station feature extraction network based on geographic information embedding, train each network module, obtain a single-station earthquake positioning system, and output the time series feature tensor, time-frequency feature tensor, and station information feature tensor;

[0062] Step 3.1: Input the preprocessed seismic waveform data into the time series feature extraction network based on multi-scale convolution and Bi-Mamba, capture the features of different scales in the time series dimension through the multi-scale convolution kernel, and combine the residual connection to retain the original waveform information; then use the bidirectional Bi-Mamba module to extract the long time series dependency features and output the time series feature tensor;

[0063] In this embodiment, the preprocessed seismic waveform data of shape (B, 3, 4000) is input into the time series feature extraction network based on multi-scale convolution and Bi-Mamba, where B is the batch size batch_size, and 4000 is the number of sampling points per channel; two convolution kernels of different sizes (such as 3 and 5) are used for parallel convolution, and the formula is as follows:

[0064] c k =ω k *x+b k ,k∈{3,5}

[0065] where ω k represents the weight of the convolution kernel, b kRepresents the bias term, and * represents the convolution operation. The shape of the feature tensor extracted by each convolution kernel is (B, C, 4000), where C is the number of convolution kernels, 64; the outputs of the two convolution kernels are combined to fuse features of different scales: x represents the input data, and k represents the size of the convolution kernel, that is, the receptive field range during the convolution operation.

[0066]

[0067] In order to retain the information of the original input data, the input data is passed through a 1×1 convolution layer to obtain the residual tensor r, and the convolution feature c fusion Added to the residual tensor r, we get the final convolution output:

[0068] y res =c fusion +r

[0069] The convolution output y res Input to the Bi-Mamba module for bidirectional timing modeling; Bi-Mamba consists of two directional flows, forward and backward;

[0070] Forward flow:

[0071] y forward =Proj(A·x forward +B)

[0072] x forward and x backward They represent the timing characteristics of the forward and backward flows input to the Bi-Mamba module, both of which are represented by y res Specifically, x forward It's y res , the input y res The forward flow is directly input, and x backward By y res Features obtained after flipping;

[0073] Backward flow:

[0074] y backward =Proj(C·x backward +D)

[0075] Finally, the output of the bidirectional stream is merged by addition:

[0076] y Bi-Mamba =y forward +y backward

[0077] The shape of the output tensor is (B, 64, 4000). The output of the Bi-Mamba module is globally pooled to aggregate the time dimension into a fixed length. The shape of the final output tensor is (B, 64). A and D are parameter matrices.

[0078] Step 3.2: Input the preprocessed seismic waveform data into the time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and output the time-frequency feature tensor;

[0079] Step 3.2.1: Use short-time Fourier transform (STFT) to convert the input waveform data into a two-dimensional time-frequency diagram to associate the time domain and frequency domain information;

[0080] Step 3.2.2: The time-frequency graph is divided into blocks through the Patch Embedding layer and input into the ConvMixer structure, and the time-frequency features are gradually extracted through Depthwise convolution and Pointwise convolution;

[0081] Step 3.2.3: Input the extracted time domain features into the multi-scale cross-axis attention module MCANet, perform multi-axis weighting on the time-frequency features, and generate a time-frequency feature tensor;

[0082] In this embodiment, the pre-processed seismic waveform data is input into a time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and is converted into a time-frequency graph through short-time Fourier transform (STFT) to form a two-dimensional feature matrix. The formula for short-time Fourier transform of each channel data is as follows:

[0083]

[0084] i is an imaginary unit, satisfying i 2 =-1, t is a time variable, representing the sampling point of the signal x(t) in time.

[0085] Where w(t-τ) is the Gaussian window function, x(t) is the waveform signal, τ and ω are time and frequency respectively, and the STFT operation converts the three-channel data into 6-channel data (real and imaginary parts are separated) with a shape of (6, 64, 64); the converted tensor is input into the ConvStem module; ConvStem uses a convolution kernel of size patch_size to divide the time-frequency image into a series of feature blocks in a block-by-block manner. The output size of each convolution block is determined by the number of filters, and the size of the output feature tensor is C×H ′ ×W ′, where C is the number of channels, H′ and W′ represent the height and width of the output feature map after the convolution operation; the embedded feature block is input into a series of convolutional mixing blocks (ConvMixerBlock), each of which contains a depthwise convolution Depthwise Conv and a pointwise convolution Pointwise Conv; then, the output of each convolutional block passes through a multi-scale cross-axis attention layer, the MCANet module, which uses a cross-axis attention mechanism to generate attention weights through three different pooling operations (average pooling, maximum pooling, and standard deviation pooling), thereby dynamically adjusting the feature responses at different positions. The formula for cross-axis attention is as follows:

[0086]

[0087] Among them, σ is the Sigmoid function, α i is a learnable weight parameter, pool i Represents three pooling operations (average pooling, maximum pooling, and standard deviation pooling); the output tensors of each module are aggregated in the spatial dimension through global pooling to generate a fixed-length feature vector with a shape of (B, 64);

[0088] Step 3.3: Input the geographical location information of the station into the station feature extraction network based on geographical information embedding, and use the fully connected layer to map the geographical information into the station information feature tensor; the geographical location information includes longitude, latitude and elevation;

[0089] In this embodiment, the shape of the feature tensor obtained is (B, 64);

[0090] Step 4: Concatenate the time series feature tensor, time-frequency feature tensor and station information feature tensor obtained in step 3 to generate a comprehensive feature tensor with a shape of (B, 192);

[0091] Step 5: Input the comprehensive feature tensor into the regression head, and pass through multiple layers of fully connected layers and Softplus activation function in sequence. The trained regression head outputs the predicted values ​​of focal depth and epicenter distance;

[0092]

[0093] In the formula represents the final regression prediction value of the model, specifically, the prediction result of earthquake source depth or epicenter distance. Z represents the comprehensive feature tensor input to the regression head, and b represents the bias parameter of the regression head, which is used to introduce additional corrections in the output. W reg is the weight parameter of the regression head, and Softplus is the activation function to ensure that the output value is non-negative;

[0094] Step 6: Conduct model training for the single-station source location model;

[0095] Step 6.1: Add multi-station collaborative loss. Each training batch contains 64 observation data. Every four observations are input from four different stations of the same earthquake event. For the four stations of the same event, the epicenter distance predicted by the single-station source location model and the station's geographical location are used to infer the source location, and compared with the actual source location, the position deviation is calculated as an additional loss and added to the total loss function for training.

[0096] Specifically, the geographical distance between the station and the predicted earthquake source is calculated by the Haversine formula:

[0097]

[0098] Among them, R is the radius of the earth, lat1 and lat2 represent the latitude of the station and the predicted source, respectively, Δlat = lat2-lat1 represents the latitude difference between the station and the source location, Δlon = lon2-lon1 represents the longitude difference between the station and the source location; lat2 is the latitude of the source, lat1 is the latitude of the station; lon2 is the longitude of the source, lon1 is the longitude of the station, and the source location is estimated by minimizing the difference between the predicted epicenter distance of the four stations and the actual location of the station. The estimated location is compared with the actual source location, and the position deviation is calculated and added as a collaborative loss to the total loss function for optimization. The final total loss function is:

[0099] Total Loss=DepthLoss+Distance Loss+Geographical Loss

[0100] Among them, Depth Loss and Distance Loss are the mean square errors of the predicted focal depth and epicenter distance, and Geographical Loss is the collaborative loss of multiple stations. This part of the loss guides the model to learn the positioning information of a single station more accurately.

[0101] Step 6.2: Dynamically adjust the learning rate and observe the test effect during model training. When the prediction error is greater than the set threshold and the loss is no longer reduced, automatically reduce the learning rate to avoid falling into the local optimum. When the model converges further and the loss begins to decrease, maintain or increase the learning rate by 0.001 to improve the convergence speed and stability of the model.

[0102] Step 7: Use the trained earthquake location system to predict the focal depth and epicenter distance. Input the earthquake waveform data collected by the station into the data preprocessing module to obtain standardized waveform data, which is input into the earthquake source location module to output the predicted results of the focal depth and epicenter distance of the station.

[0103] The single-station earthquake location method combining multi-source characteristics and coordinated losses described in the present invention is used to conduct experiments using data provided by the Italian strong earthquake dataset INSTANCE. The dataset contains 50,000 earthquake event data with focal depths of 0-72km, epicentral distances of 0-112km, and magnitudes below 6.5, of which 40,000 are used for training and 10,000 are used for testing. The experimental results show that the method achieves a mean absolute error (MAE) of 3.01km for focal depth estimation and a MAE of 2.57km for epicentral distance estimation on the test set. The results show that the method of the present invention can effectively determine the focal depth and epicentral distance within an acceptable error range using only single station information. The method of the present invention is not only simple to deploy and efficient in calculation, but also solves the problem that earthquake location can only be performed by relying on the observation data of a single station due to sparse distribution of seismic stations or terrain restrictions in many areas. It has good practicality in the application of focal depth and epicentral distance prediction, and meets the positioning requirements in earthquake monitoring and emergency response.

[0104] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.

Claims

1. A single-station earthquake location system combining multi-source characteristics and coordinated loss, characterized in that: include: Data preprocessing module, model training module, single-station source location model; The data preprocessing module is used to preprocess the collected original seismic waveform data to obtain the preprocessed seismic waveform data, that is, the seismic three-channel waveform data; The model training module inputs the three-channel earthquake waveform data into the single-station source location model, trains the single-station source location model through collaborative loss, and dynamically adjusts the learning rate during the training process to obtain the trained single-station source location model; The single-station earthquake source location model is a three-branch deep learning network. The earthquake waveform data collected by the station is input into a data preprocessing module to obtain the preprocessed earthquake waveform data, which is input into the trained single-station earthquake source location model to output the prediction results of the earthquake source depth and epicenter distance.

2. A single-station earthquake location system combining multi-source characteristics and coordinated loss according to claim 1, characterized in that: The collaborative loss is specifically as follows: on the basis of the mean square error, the collaborative loss of multiple stations is added. First, the three-channel earthquake waveform data is adjusted so that each training batch contains multiple stations of the same earthquake event. Then, when calculating the loss, the epicenter distance of the station predicted by the model and the geographical location information of the station are used to estimate the position of a source through a geometric method. Finally, the spherical distance between the estimated source and the actual source is calculated as the collaborative loss auxiliary model training.

3. A single-station earthquake location system combining multi-source characteristics and coordinated loss according to claim 1, characterized in that: The single-station source location model specifically includes: a time series feature extraction network based on multi-scale convolution and Bi-Mamba, a time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and a station feature extraction network based on geographic information embedding; The time series feature extraction network based on multi-scale convolution and Bi-Mamba uses a multi-scale convolution kernel to extract multi-scale features from the input three-channel seismic waveform data, and enhances the integrity of the features through residual connections; then the Bi-Mamba module is used to capture the long-term dependency of the waveform data, obtain the time series feature tensor, and realize preliminary time series modeling; The time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention first generates a time-frequency graph through short-time Fourier transform STFT of the input three-channel seismic waveform data, and then divides the time-frequency graph into blocks using the Patch Embedding layer and inputs it into the ConvMixer network, gradually extracts multi-scale spatial and frequency features, and finally generates a time-frequency feature tensor; the time-frequency feature tensor passes through the MCANet module, and weights the key features at different scales based on multi-scale cross-axis attention to obtain a time-frequency feature tensor; The station feature extraction network based on geographic information embedding embeds the geographic location information of the station, converts it into a feature tensor through a fully connected layer, and obtains a station information feature tensor; The time series feature tensors, time-frequency feature tensors and station information feature tensors output by the above three networks are concatenated to generate a comprehensive feature tensor; the comprehensive feature tensor passes through the regression head of multiple layers of fully connected layers, and the final predicted values ​​of focal depth and epicenter distance are output with the Softplus activation function.

4. A single-station earthquake positioning method combining multi-source characteristics and coordinated loss, based on the single-station earthquake positioning system combining multi-source characteristics and coordinated loss according to claim 1, characterized in that: The following steps are involved: Step 1: Query the earthquake catalog information to determine the arrival time of the three-component earthquake P wave; after determining the first arrival time of the P wave, intercept the data from a second before to b seconds after the starting point of each component; Step 2: Input the obtained three-component waveform data into the data preprocessing module, remove the seismic waveform data containing discontinuities and the signal-to-noise ratio below the set threshold, divide the seismic waveform data by the sensitivity of the seismometer, and then filter, mean detrend and normalize the waveform; Step 3: Input the preprocessed seismic waveform data into the time series feature extraction network based on multi-scale convolution and Bi-Mamba, the time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and the station feature extraction network based on geographic information embedding, train each network module, obtain a single-station earthquake positioning system, and output the time series feature tensor, time-frequency feature tensor, and station information feature tensor; Step 4: Concatenate the time series feature tensor, time-frequency feature tensor and station information feature tensor obtained in step 3 to generate a comprehensive feature tensor; Step 5: Input the comprehensive feature tensor into the regression head, and pass through multiple layers of fully connected layers and Softplus activation function in sequence. The trained regression head outputs the predicted values ​​of focal depth and epicenter distance; Step 6: Conduct model training for the single-station source location model; Step 7: Use the trained earthquake location system to predict the focal depth and epicenter distance. Input the earthquake waveform data collected by the station into the data preprocessing module to obtain standardized waveform data, which is input into the earthquake source location module to output the predicted results of the focal depth and epicenter distance of the station.

5. A single-station earthquake location method combining multi-source characteristics and coordinated loss according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1: Input the preprocessed seismic waveform data into the time series feature extraction network based on multi-scale convolution and Bi-Mamba, capture the features of different scales in the time series dimension through the multi-scale convolution kernel, and combine the residual connection to retain the original waveform information; then use the bidirectional Bi-Mamba module to extract the long time series dependency features and output the time series feature tensor; Step 3.2: Input the preprocessed seismic waveform data into the time-frequency feature extraction network based on short-time Fourier transform and cross-axis attention, and output the time-frequency feature tensor; Step 3.2.1: Use short-time Fourier transform (STFT) to convert the input waveform data into a two-dimensional time-frequency diagram to associate the time domain and frequency domain information; Step 3.2.2: The time-frequency graph is divided into blocks through the Patch Embedding layer and input into the ConvMixer structure, and the time-frequency features are gradually extracted through Depthwise convolution and Pointwise convolution; Step 3.2.3: Input the extracted time domain features into the multi-scale cross-axis attention module MCANet, perform multi-axis weighting on the time-frequency features, and generate a time-frequency feature tensor; Step 3.3: Input the geographic location information of the station into the station feature extraction network based on geographic information embedding, and use the fully connected layer to map the geographic information into a station information feature tensor; the geographic location information includes longitude, latitude and elevation.

6. A single-station earthquake location method combining multi-source characteristics and coordinated loss according to claim 1, characterized in that: The predicted values ​​of focal depth and epicenter distance in step 5 are as follows In the formula represents the final regression prediction value of the model, Z represents the comprehensive feature tensor input to the regression head, b represents the bias parameter of the regression head, and W reg is the weight parameter of the regression head, and Softplus is the activation function to ensure that the output value is non-negative.

7. A single-station earthquake location method combining multi-source characteristics and coordinated loss according to claim 1, characterized in that: The step 6 comprises the following steps: Step 6.1: Add multi-station collaborative loss. Each training batch contains 64 observation data. Every four observations are input from four different stations of the same earthquake event. For the four stations of the same event, the epicenter distance predicted by the single-station source location model and the station's geographical location are used to infer the source location, and compared with the actual source location, the position deviation is calculated as an additional loss and added to the total loss function for training. Specifically, the geographical distance between the station and the predicted earthquake source is calculated by the Haversine formula: Among them, R is the radius of the earth, lat1 and lat2 represent the latitude of the station and the predicted source, respectively, Δlat = lat2-lat1 represents the latitude difference between the station and the source location, Δlon = lon2-lon1 represents the longitude difference between the station and the source location; lat2 is the latitude of the source, lat1 is the latitude of the station; lon2 is the longitude of the source, lon1 is the longitude of the station, and the source location is estimated by minimizing the difference between the predicted epicenter distance of the four stations and the actual location of the station. The estimated location is compared with the actual source location, and the position deviation is calculated and added as a collaborative loss to the total loss function for optimization. The final total loss function is: TotalLoss=DepthLoss+Distance Loss+GeographicalLoss Among them, Depth Loss and Distance Loss are the mean square errors of the predicted focal depth and epicenter distance, and Geographical Loss is the coordination loss of multiple stations; Step 6.2: Dynamically adjust the learning rate and observe the test effect during model training. When the prediction error is greater than the set threshold and the loss no longer decreases, automatically reduce the learning rate. When the model converges further and the loss begins to decrease, maintain or increase the learning rate by 0.001.

Images