Seismic interference signal rejection model construction method, seismic interference signal rejection method and seismic interference signal rejection system

Abstract

The invention discloses a seismic interference signal elimination model construction method, elimination method and system, and relates to the field of geophysical data processing, and the method comprises the steps: obtaining historical waveform data of a trigger event from a historical seismic trigger database; generating a training sample set based on the labeling result of the historical waveform data, inputting the training sample set into a long short-term memory network, and performing time sequence feature extraction on the historical waveform data by using the long short-term memory network to obtain seismic signal time sequence features; based on the seismic signal time sequence features, calculating seismic signal aggregation features by using a feature weighted aggregation algorithm; calculating a prediction category of the historical waveform data through a classification decision algorithm based on the seismic signal aggregation features; and finally, based on the historical waveform data, the labeling category and the prediction category, performing iterative training on the long-short-term memory network to obtain a trained seismic interference signal elimination model. Interference signals can be effectively eliminated, and the seismic data processing reliability is improved.

Description

Technical Field

[0001] This invention relates to the field of geophysical data processing technology, specifically to a method for constructing a model for removing seismic interference signals, a removal method, and a system. Background Technology

[0002] Currently, in earthquake monitoring and automated cataloging systems, the identification and classification of event signals are crucial for ensuring data quality and accurate early warning. Traditional earthquake event discrimination methods typically rely on a combination of manual feature extraction and traditional machine learning algorithms. Specifically, researchers manually extract a series of feature parameters from earthquake waveforms, such as peak frequency, spectral width, P / S wave energy ratio, body wave magnitude to surface wave magnitude ratio, envelope morphology, and polarization characteristics. Then, they use classification models such as Support Vector Machines (SVM), Random Forests, or shallow neural networks to distinguish earthquake events from non-earthquake interference signals.

[0003] The above methods have the following objective disadvantages: (1) Such algorithms are highly dependent on manual feature engineering. Different types of interference signals (such as traffic vibration, blasting, instrument pulses, etc.) have significant differences in waveform features, and fixed manual features are often difficult to cover all interference modes, resulting in weak generalization ability of the model. (2) The feature extraction process requires a lot of knowledge in the field of seismology and manual participation, and the design and maintenance costs are high, which is not conducive to its promotion and application in large-scale automated systems. (3) Traditional methods classify based on statistical features, which is difficult to fully reflect the dynamic evolution characteristics of waveform signals in the time dimension (such as the successive appearance of P waves, S waves and coma waves), thus limiting the model's recognition accuracy for complex signals.

[0004] In summary, existing methods for identifying earthquake early warning events typically rely on manual feature extraction combined with machine learning algorithms, which makes it difficult to effectively capture the temporal dynamic features of earthquake waveforms and limits the accuracy of models in recognizing complex signals. Summary of the Invention

[0005] To address the problem that existing earthquake early warning technologies struggle to effectively capture the temporal dynamics of earthquake waveforms, thus limiting the accuracy of models in recognizing complex signals.

[0006] In a first aspect, the present invention proposes a method for constructing a seismic interference signal elimination model, comprising: Historical waveform data of triggering events are obtained from a historical earthquake triggering database; a training sample set is generated based on the annotation results of the historical waveform data, the training sample set including the historical waveform data and its corresponding annotation categories; The training sample set is input into a long short-term memory network, and the long short-term memory network is used to extract the temporal features of the historical waveform data to obtain the temporal features of the seismic signal. Based on the temporal features of the seismic signal, a feature weighted aggregation algorithm is used to calculate the aggregated features of the seismic signal. Based on the aggregated features of the seismic signal, a classification decision algorithm is used to calculate the predicted category of the historical waveform data. Based on the historical waveform data, the labeled categories, and the predicted categories, the long short-term memory network is iteratively trained to obtain a trained earthquake interference signal removal model.

[0007] Preferably, historical waveform data of the triggering event is obtained from a historical earthquake triggering database, including: Data on multiple triggered earthquake events were obtained from the historical earthquake triggering database of the historical seismic network. For each triggered earthquake event, the historical waveform data of the event is extracted using a fixed-length time window based on the triggering time. The historical waveform data includes vertical, north-south, and east-west components.

[0008] Preferably, generating the training sample set based on the annotation results of the historical waveform data includes: The historical waveform data is reviewed and labeled, and based on the review and labeling results, the corresponding earthquake event category label is labeled for each time window data to generate a training sample set; Specifically, when the earthquake event category label value is 1, it indicates that an earthquake event has occurred; When the earthquake event category label value is 0, it indicates that an interference event has occurred.

[0009] Preferably, before inputting the training sample set into the Long Short-Term Memory network, the method further includes: The mean and linear detrending processes are performed on all components of each historical waveform data in the training sample set, and a bandpass filter is applied for noise reduction to obtain each historical waveform data after noise reduction. All components in each historical waveform data after noise reduction are normalized.

[0010] Preferably, the step of using the Long Short-Term Memory network to extract time-series features from the historical waveform data to obtain the seismic signal time-series features includes: A temporal feature extraction module employing a long short-term memory network model extracts the seismic signal temporal features of the historical waveform data in the time dimension; the seismic signal temporal features are used to identify seismic signals.

[0011] Preferably, the step of calculating the aggregated features of the seismic signal using a feature-weighted aggregation algorithm based on the time-series features of the seismic signal includes: Based on the temporal characteristics of the earthquake signal, an attention mechanism algorithm is used to calculate the attention weight at each moment in the temporal characteristics of the earthquake signal. Based on the attention weights and the temporal features of the seismic signal at each time step, the feature weighting aggregation module of the long short-term memory network model is used to calculate the aggregated features of the seismic signal.

[0012] Furthermore, the function expression of the feature weighted aggregation module is:

[0013] in, Let be the attention weight at time t; Z represents the hidden state at time t; Z represents the seismic signal aggregation feature.

[0014] Preferably, the step of calculating the predicted category of the historical waveform data based on the seismic signal aggregation features using a classification decision algorithm includes: The aggregated seismic signal features are processed by nonlinear activation through the fully connected layer of the long short-term memory network model to obtain seismic feature spatial data. Based on the aforementioned seismic feature spatial data, the classification decision module of the Long Short-Term Memory Network model is used to determine the predicted category of the historical waveform data.

[0015] Secondly, this invention proposes a seismic interference signal removal model construction system, comprising: Data acquisition module: used to obtain historical waveform data of triggering events from the historical earthquake triggering database; Annotation module: used to generate a training sample set based on the annotation results of the historical waveform data, the training sample set including the historical waveform data and its corresponding annotation categories; Temporal feature extraction module: used to input the training sample set into the long short-term memory network, and use the long short-term memory network to extract temporal features from the historical waveform data to obtain the temporal features of the seismic signal; Feature aggregation module: used to calculate the aggregated features of the seismic signal based on the time-series features of the seismic signal using a feature-weighted aggregation algorithm; Classification and decision module: used to obtain the predicted category of the calculated historical waveform data based on the aggregated features of the seismic signal through a classification and decision algorithm; Training module: used to iteratively train the long short-term memory network based on the historical waveform data, the labeled categories, and the predicted categories to obtain a trained earthquake interference signal removal model.

[0016] Preferably, the data acquisition module is specifically used to acquire multiple triggered earthquake event record data from the historical earthquake triggering database of the historical seismic network; For each triggered earthquake event, the historical waveform data of the event is extracted using a fixed-length time window based on the triggering time. The historical waveform data includes vertical, north-south, and east-west components.

[0017] Preferably, the annotation module is specifically used to review and annotate the historical waveform data, and to annotate the corresponding earthquake event category label for each time window data according to the review results, thereby generating a training sample set; Specifically, when the earthquake event category label value is 1, it indicates that an earthquake event has occurred; When the earthquake event category label value is 0, it indicates that an interference event has occurred.

[0018] Preferably, the system further includes: The data preprocessing module is used to perform mean and linear detrending processing on all components of each historical waveform data in the training sample set, and to apply a bandpass filter for noise reduction. All components in each historical waveform data after noise reduction are normalized.

[0019] Preferably, the temporal feature extraction module is specifically used to extract the temporal features of the seismic signal in the time dimension using a long short-term memory network model temporal feature extraction module; the seismic signal temporal features are used to identify the seismic signal.

[0020] Preferably, the feature aggregation module is specifically used to calculate the attention weight at each moment in the seismic signal time series features based on the seismic signal time series features using an attention mechanism algorithm; Based on the attention weights and the temporal features of the seismic signal, a feature weighted aggregation module is used to calculate the aggregated features of the seismic signal.

[0021] Furthermore, the function expression of the feature weighted aggregation module is:

[0022] in, Let be the attention weight at time t; Z represents the hidden state at time t; Z represents the seismic signal aggregation feature.

[0023] Preferably, the classification decision module is specifically used to perform nonlinear activation processing on the aggregated features of the seismic signal through a fully connected layer to obtain seismic feature spatial data; Based on the aforementioned seismic feature spatial data, the predicted category of the historical waveform data is calculated using a classification decision module according to the classification type of the seismic signal.

[0024] Thirdly, the present invention proposes an interference signal removal method based on an earthquake interference signal removal model, comprising: When a seismic station issues a trigger event report, the current waveform data is obtained from the trigger event report; The earthquake waveform data is processed by removing interference signals using a pre-built earthquake interference signal removal model to obtain the event category corresponding to the current waveform data. Based on the event category, determine whether the triggered event report is a real earthquake event; if so, trigger an earthquake early warning. The earthquake interference signal removal model is constructed using a method for constructing an earthquake interference signal removal model as described in any one of claims 1 to 8.

[0025] Preferably, determining whether the triggering event report is a real earthquake event based on the event category includes: When the trigger probability of the event category is greater than a preset threshold, the trigger event report is determined to be a real earthquake event; When the trigger probability of the event category is not greater than a preset threshold, the trigger event report is determined to be an interference event.

[0026] Fourthly, this invention proposes an interference signal removal system based on an earthquake interference signal removal model, comprising: Data acquisition module: When a seismic station issues a trigger event report, it acquires the current waveform data from the trigger event report; Interference signal removal module: Uses a pre-built seismic interference signal removal model to remove interference signals from the seismic waveform data to obtain the event category corresponding to the current waveform data; Earthquake early warning module; used to determine whether the triggered event report is a real earthquake event based on the event category, and if so, to trigger an earthquake early warning; The earthquake interference signal elimination model is constructed using a method for constructing an earthquake interference signal elimination model as described in any one of claims 1 to 8.

[0027] Preferably, the earthquake early warning module is specifically used to: when the trigger probability of the event category is greater than a preset threshold, determine that the triggering event report is a real earthquake event, and then trigger an earthquake early warning; When the trigger probability of the event category is not greater than a preset threshold, the trigger event report is determined to be an interference event and no earthquake early warning is triggered.

[0028] Furthermore, this application also provides a computing device, comprising: one or more processors; A processor is used to execute one or more programs; When the one or more programs are executed by the one or more processors, they implement a method for constructing an earthquake interference signal removal model as described above, and a method for removing interference signals based on an earthquake interference signal removal model as described above.

[0029] In another aspect, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements a method for constructing an earthquake interference signal removal model as described in any of the preceding claims, and a method for removing interference signals based on an earthquake interference signal removal model as described in any of the preceding claims.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method, method, and system for constructing an earthquake interference signal removal model. The method involves obtaining historical waveform data of triggering events from a historical earthquake triggering database; generating a training sample set based on the annotation results of the historical waveform data; inputting the training sample set into a Long Short-Term Memory (LSTM) network; using the LSM network to extract temporal features from the historical waveform data to obtain temporal features of the earthquake signal; calculating aggregated features of the earthquake signal using a feature weighted aggregation algorithm based on the temporal features of the earthquake signal; calculating the predicted category of the historical waveform data using a classification decision algorithm based on the aggregated features of the earthquake signal; and finally, iteratively training the LSM network based on the historical waveform data, the annotation categories, and the predicted categories to obtain a trained earthquake interference signal removal model. This invention utilizes a long short-term memory network to perform end-to-end intelligent analysis and discrimination of detected candidate seismic event waveforms. By automatically learning the inherent differences between seismic signals and interference signals in time series patterns, it achieves high-precision automatic classification of candidate event waveforms. The seismic interference signal removal model constructed using this invention can automatically learn features from the original seismic waveforms, possess time series modeling capabilities, and improve classification accuracy, thereby effectively removing interference signals, reducing manual intervention, and enhancing the automation and reliability of seismic data processing. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the process for constructing a seismic interference signal removal model according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the LSTM network structure in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of a seismic interference signal removal model construction system according to Embodiment 2 of the present invention; Figure 4 This is a schematic diagram of the electronic device in Embodiment 5 of the present invention. Detailed Implementation

[0032] Example 1: This invention provides a method for constructing a seismic interference signal removal model, such as... Figure 1 As shown, it includes the following steps: Step S1: Obtain historical waveform data of the triggering event from the historical earthquake triggering database; Step S2: Generate a training sample set based on the annotation results of the historical waveform data. The training sample set includes the historical waveform data and its corresponding annotation categories. Step S3: Input the training sample set into the Long Short-Term Memory network, and use the Long Short-Term Memory network to extract the temporal features of the historical waveform data to obtain the temporal features of the seismic signal; Step S4: Based on the time-series characteristics of the seismic signal, the feature-weighted aggregation algorithm is used to calculate the aggregated features of the seismic signal; Step S5: Based on the aggregation features of the seismic signal, calculate the predicted category of the historical waveform data using a classification decision algorithm; Step S6: Based on the historical waveform data, the labeled categories, and the predicted categories, the long short-term memory network is iteratively trained to obtain a trained earthquake interference signal removal model.

[0033] Specifically, in step S1 of this embodiment of the invention, obtaining historical waveform data of the triggering event from the historical earthquake triggering database includes: Multiple triggered earthquake event records are obtained from the historical earthquake triggering database of the historical earthquake network; for each triggered earthquake event record, the historical waveform data of the triggering event is extracted using a fixed-length time window based on the triggering time; wherein, the historical waveform data includes vertical components, north-south components and east-west components.

[0034] It should be further explained that the construction and labeling of the classification training sample library in step S1 is the foundation of the entire method, which aims to provide a large-scale, class-balanced and accurately labeled dataset for model training.

[0035] First, tens of thousands, even millions, of triggered event records are collected from the historical seismic network's trigger event database. For each triggered event record, a fixed-length time window is extracted based on its trigger time. This window should be long enough to completely encompass the main part of a typical near-earthquake event (i.e., historical waveform data). This means selecting a fixed-duration waveform with distinct seismic characteristics from the earthquake's trigger waveform data. Different input lengths will result in different characteristics, and the extraction can be tailored to requirements such as timeliness. For example, a 10-second segment forward and a 50-second segment backward could be extracted, for a total duration of 60 seconds. The extracted historical waveform data should contain all three components (e.g., vertical Z-axis, north-south N-axis, and east-west E-axis).

[0036] Specifically, in step S2 of this embodiment of the invention, generating a training sample set based on the annotation results of the historical waveform data includes: The historical waveform data is reviewed and labeled. Based on the review and labeling results, the corresponding earthquake event category label is labeled for each time window data to generate a training sample set. Among them, when the earthquake event category label value is 1, it indicates that an earthquake event has occurred; when the earthquake event category label value is 0, it indicates that an interference event has occurred.

[0037] Further explanation is needed: in step S2, experienced earthquake analysis experts meticulously review and label each of the extracted event waveform windows (i.e., historical waveform data). Each window is assigned a unique category label. In a basic binary classification scenario, the label can be set to "0" for "earthquake event" and "1" for "interference event." For more refined discrimination capabilities, a multi-classification scheme can also be used. All labeled event windows and their corresponding category labels together constitute the dataset used for model training, validation, and testing.

[0038] Specifically, to ensure stable and efficient model learning, each extracted event waveform window needs to undergo consistent preprocessing before being fed into the network. In step S3 of this embodiment, before inputting the training sample set into the Long Short-Term Memory network, the following steps are also included: For each historical waveform data in the training sample set, all components are subjected to mean and linear detrending processing, and a bandpass filter is applied for noise reduction to obtain each historical waveform data after noise reduction. Then, all components in each historical waveform data after noise reduction are normalized. This is the preprocessing of the event waveform data.

[0039] The preprocessing procedure, which requires further explanation, includes: Detrending and Filtering: For the historical waveform data of each component, mean removal and linear detrending are first performed to eliminate baseline drift. Subsequently, a mild bandpass filter (e.g., 1-20Hz) is applied to filter out low-frequency and high-frequency noise that is irrelevant to the target seismic signal.

[0040] Normalization: The amplitude of the entire event window (containing all components) is normalized. A preferred implementation is to calculate the maximum absolute value of all sampling points within the window across all components, and then use this value to divide all data points within the window, so that the amplitude of all waveforms is scaled to the interval [-1, 1]. This window-level normalization preserves the relative amplitude relationships between different components, which is an important physical characteristic.

[0041] In this embodiment of the invention, a Long Short-Term Memory (LSTM) network event classification model is constructed and trained, which is a deep classification network for processing multivariate time series. The core of this network is the LSTM network, which can capture the dynamic characteristics of waveform evolution over time. This is crucial for distinguishing earthquakes with specific development patterns (such as the sequential appearance of P-waves, S-waves, and codas) from interferences with different patterns.

[0042] Specifically, step S3 of this embodiment of the invention, which involves using the long short-term memory network to extract time-series features from the historical waveform data to obtain the time-series features of the seismic signal, includes: A temporal feature extraction module employing a long short-term memory network model extracts the seismic signal temporal features of the historical waveform data in the time dimension; the seismic signal temporal features are used to identify seismic signals.

[0043] Specifically, in step S4 of this embodiment of the invention, based on the time-series characteristics of the seismic signal, a feature-weighted aggregation algorithm is used to calculate the aggregated features of the seismic signal, including: Based on the temporal characteristics of the earthquake signal, an attention mechanism algorithm is used to calculate the attention weight at each moment in the temporal characteristics of the earthquake signal; based on the attention weight at each moment and the temporal characteristics of the earthquake signal, the feature weighting aggregation module of the long short-term memory network model is used to calculate the aggregated features of the earthquake signal.

[0044] Specifically, in step S5 of this embodiment of the invention, calculating the predicted category of the historical waveform data based on the seismic signal aggregation features using a classification decision algorithm includes: The aggregated seismic signal features are processed by nonlinear activation through the fully connected layer of the long short-term memory network model to obtain seismic feature spatial data. Based on the aforementioned seismic feature spatial data, the classification decision module of the Long Short-Term Memory Network model is used to determine the predicted category of the historical waveform data.

[0045] It should be further explained that the network architecture of the Long Short-Term Memory network event classification model in this embodiment of the invention is as follows: Figure 2 As shown: Input layer: Receives preprocessed tensors, such as [time step, number of components] (60 seconds, 100Hz sampling is [6000, 3]).

[0046] Temporal Feature Extraction Module (corresponding to step S3): In this module, the input seismic waveform data (i.e., historical waveform data in the training sample set, usually multi-component time series, such as waveforms in the Z, N, and E directions) is first processed by a Long Short-Term Memory (LSTM) network to extract its dynamic features in the time dimension (i.e., seismic signal temporal features). The LSTM network can capture the pattern of energy change over time in the waveform, thereby identifying key features of the seismic signal, such as the rapid initial motion of the P-wave, the strong peak of the S-wave, and the attenuation trend of the coma. The LSTM network extracts not only seismic signal features but also interference signal features. By learning the evolution patterns of different signals in the time dimension, the model can distinguish between seismic and interference signals.

[0047] Feature aggregation module (corresponding to step S4): To further enhance the model's adaptive judgment of the importance of different time segments, the system introduces an attention mechanism. This mechanism automatically calculates weights based on the relevance of the hidden state at each time step, enabling the model to focus on the most discriminative parts of the seismic signal, such as the initial segment of the P-wave or the energy peak of the S-wave, while assigning lower weights to noisy segments.

[0048] Specifically, the attention mechanism adjusts the weights by calculating the "importance" at each moment, where... Let represent the attention weight at time t, which is determined by scoring the hidden state at each time step. The formula for calculating this weight is:

[0049] In this formula, This represents the score for the hidden state at time t. The numerator represents the score of the hidden state at time i; the numerator emphasizes the weight of important times through an exponential function, while the denominator normalizes the hidden states at all times to ensure that the sum of all weights is 1; T refers to the total number of time steps in the input sequence.

[0050] It's important to further clarify that in seismic signal analysis, a time step typically refers to each data point obtained from sampling. For example, assuming seismic waveform data is acquired at a sampling rate of 100Hz, each time step corresponds to each sampling point. In a Long Short-Term Memory (LSTM) network, the input data for each time step (e.g., seismic waveform data from the Z, N, and E directions) is fed into the network, and its hidden states and cell states are progressively updated by the internal units of the LSTM. The output of each time step is the hidden state at that moment, representing the temporal characteristics of the input seismic signal and its relationship with previous time steps.

[0051] Next, the feature weighted aggregation process uses the calculated attention weights. To weight the hidden state at each time step This ultimately generates a high-dimensional event feature vector (i.e., aggregated seismic signal features). The weighted aggregation formula is as follows:

[0052] in, Let be the attention weight at time t; Z represents the hidden state at time t; Z represents the seismic signal aggregation feature.

[0053] The aggregated features of the seismic signal are represented by the hidden state ht at each time step, which reflects the temporal characteristics of the seismic signal in the input seismic waveform. The high-dimensional event feature vector z is a weighted sum of the hidden states at all time steps after being weighted by an attention mechanism; it represents the comprehensive features of the entire seismic event.

[0054] Classification decision module (classification head) (corresponding to step S5): The feature vector (i.e., the aggregated features of the seismic signal) is first processed through a fully connected layer (ReLU activation) to map it to a new feature space for classification decisions. Subsequently, the output is passed through a Softmax or Sigmoid layer for probability calculation. Softmax is used for multi-class classification tasks (such as earthquakes, blasting, traffic noise, etc.). The formula for Softmax is:

[0055] In this formula, represent Corresponding to category The original score, This represents the sum of the indices of the scores for all categories. C represents the total number of categories, indicating the number of distinct categories in the output layer.

[0056] For binary classification tasks, the Sigmoid function is used, and its formula is:

[0057] Finally, based on the set threshold (e.g., 0.9), the model will determine whether to trigger an early warning. If the probability of an earthquake is greater than the threshold, it will be determined as an earthquake event; otherwise, it will be considered an interference signal.

[0058] During model training, labeled seismic event data is first preprocessed, including detrending, bandpass filtering, and normalization. The data is then input into an LSTM-based feature extraction module to extract temporal features and capture key seismic signal characteristics such as P-waves and S-waves. Next, an attention mechanism is used to weight the importance of different time segments, generating a high-dimensional event feature vector. This feature vector is processed through a fully connected layer, and the output is used to generate class probabilities through a Softmax or Sigmoid layer. Training employs classification cross-entropy as the loss function, and the network weights are continuously updated using the Adam optimizer and backpropagation algorithm. To prevent overfitting, early stopping and learning rate decay strategies are used, and model performance is monitored on an independent validation set. Finally, the model that performs best on the validation set is selected as the classifier.

[0059] This invention provides a method for constructing an earthquake interference signal removal model. The method involves obtaining historical waveform data of triggering events from a historical earthquake triggering database; generating a training sample set based on the annotation results of the historical waveform data; inputting the training sample set into a Long Short-Term Memory (LSTM) network; using the LSM network to extract temporal features from the historical waveform data to obtain the temporal features of the earthquake signal; calculating the aggregated features of the earthquake signal using a feature weighted aggregation algorithm based on the temporal features; calculating the predicted category of the historical waveform data using a classification decision algorithm based on the aggregated features; and finally, iteratively training the LSM network based on the historical waveform data, the annotation categories, and the predicted categories to obtain a trained earthquake interference signal removal model. This invention utilizes a deep learning model to automatically learn the inherent differences between earthquake signals and interference signals in time-series patterns, achieving high-precision automatic classification of candidate event waveforms, thereby effectively removing interference signals, reducing manual intervention, and improving the automation and reliability of earthquake data processing. This intelligent earthquake interference removal method employs a long short-term memory network in a recurrent neural network to perform end-to-end intelligent analysis and discrimination of detected candidate seismic event waveforms. It automatically learns features from the original seismic waveforms, possesses time-series modeling capabilities, and improves classification accuracy.

[0060] Example 2: Based on the same inventive concept, this invention also provides a system for constructing a seismic interference signal removal model, such as... Figure 3 As shown, it includes: Data acquisition module: used to obtain historical waveform data of triggering events from the historical earthquake triggering database; Annotation module: used to generate a training sample set based on the annotation results of the historical waveform data, the training sample set including the historical waveform data and its corresponding annotation categories; Temporal feature extraction module: used to input the training sample set into the long short-term memory network, and use the long short-term memory network to extract temporal features from the historical waveform data to obtain the temporal features of the seismic signal; Feature aggregation module: used to calculate the aggregated features of the seismic signal based on the time-series features of the seismic signal using a feature-weighted aggregation algorithm; Classification and decision module: used to obtain the predicted category of the calculated historical waveform data based on the aggregated features of the seismic signal through a classification and decision algorithm; Training module: used to iteratively train the long short-term memory network based on the historical waveform data, the labeled categories, and the predicted categories to obtain a trained earthquake interference signal removal model.

[0061] Specifically, in this embodiment of the invention, the data acquisition module is specifically used to acquire multiple triggered earthquake event record data from the historical earthquake triggering database of the historical seismic network; For each triggered earthquake event, the historical waveform data of the event is extracted using a fixed-length time window based on the triggering time. The historical waveform data includes vertical, north-south, and east-west components.

[0062] Specifically, in this embodiment of the invention, the annotation module is used to review and annotate the historical waveform data, and to annotate the corresponding earthquake event category label for each time window data according to the review results, thereby generating a training sample set; Specifically, when the earthquake event category label value is 1, it indicates that an earthquake event has occurred; When the earthquake event category label value is 0, it indicates that an interference event has occurred.

[0063] Specifically, the system described in this embodiment of the invention further includes: The data preprocessing module is used to perform mean and linear detrending processing on all components of each historical waveform data in the training sample set, and to apply a bandpass filter for noise reduction. All components in each historical waveform data after noise reduction are normalized.

[0064] Specifically, in this embodiment of the invention, the temporal feature extraction module is used to extract the temporal features of the seismic signal in the time dimension using a long short-term memory network model temporal feature extraction module; the seismic signal temporal features are used to identify the seismic signal.

[0065] Specifically, in this embodiment of the invention, the feature aggregation module is specifically used to calculate the attention weight at each moment in the seismic signal time series features based on the seismic signal time series features using an attention mechanism algorithm; Based on the attention weights and the temporal features of the seismic signal, a feature weighted aggregation module is used to calculate the aggregated features of the seismic signal.

[0066] It should be further explained that the function expression of the feature weighting aggregation module in this embodiment of the invention is as follows:

[0067] in, Let be the attention weight at time t; Z represents the hidden state at time t; Z represents the seismic signal aggregation feature.

[0068] Specifically, in this embodiment of the invention, the classification decision module is used to perform nonlinear activation processing on the aggregated features of the seismic signal through a fully connected layer to obtain seismic feature spatial data; Based on the aforementioned seismic feature spatial data, the predicted category of the historical waveform data is calculated using a classification decision module according to the classification type of the seismic signal.

[0069] Example 3: Based on the same inventive concept, the present invention also provides an interference signal removal method based on an earthquake interference signal removal model, including: When a seismic station issues a trigger event report, the current waveform data is obtained from the trigger event report; The earthquake waveform data is processed by removing interference signals using a pre-built earthquake interference signal removal model to obtain the event category corresponding to the current waveform data. Based on the event category, determine whether the triggered event report is a real earthquake event; if so, trigger an earthquake early warning. The earthquake interference signal elimination model is constructed using a method for constructing an earthquake interference signal elimination model as described in Example 1.

[0070] Specifically, in this embodiment of the invention, determining whether the triggered event report is a real earthquake event based on the event category includes: When the trigger probability of the event category is greater than a preset threshold, the triggered event report is determined to be a real earthquake event; when the trigger probability of the event category is not greater than the preset threshold, the triggered event report is determined to be a disturbance event.

[0071] It should be further explained that the trained seismic interference signal removal model is deployed into the real-time seismic data processing workflow, immediately following the initial event triggering module. When any station's triggering algorithm reports a candidate event, the system immediately extracts the corresponding waveform data window according to the aforementioned specifications, performs the same preprocessing procedure, and then inputs it into the loaded LSTM classification model. The model quickly outputs the classification result of the event (i.e., its probability of belonging to "earthquake" or "interference"). Based on a preset confidence threshold (e.g., when the probability of "earthquake" is greater than 0.90), the system automatically makes a decision: events determined to be real earthquakes will be passed to subsequent fine-grained processing stages such as phase picking, location, and magnitude measurement, and can be used to trigger early warnings; events determined to be interference will be marked, archived, or directly discarded, no longer consuming the analyst's time or interfering with subsequent processes.

[0072] During earthquake monitoring and early warning, seismic waveform signals acquired by seismic monitoring stations often contain a large amount of non-seismic interference, such as traffic vibrations, blasting, wind and waves, and instrument pulses. These interference signals are similar in waveform to real earthquake events, easily leading to misjudgments. Existing seismic signal classification methods mostly employ manual feature extraction combined with machine learning algorithms, extracting parameters such as peak frequency, energy ratio, envelope features, and polarization features, and then using support vector machines (SVM) or random forests for classification. However, these methods rely on manual feature design, have limited generalization ability, and struggle to effectively capture the temporal dynamic characteristics of seismic waveforms.

[0073] This invention is an intelligent seismic interference removal method that can automatically learn features directly from raw seismic waveforms, possesses time-series modeling capabilities, and improves classification accuracy. This invention provides an end-to-end data-driven solution that utilizes a deep learning model to automatically learn the inherent differences in time-series patterns between seismic signals and interference signals, achieving high-precision automatic classification of candidate event waveforms. This effectively removes interference signals, reduces manual intervention, and improves the automation and reliability of seismic data processing.

[0074] This invention discloses a method for classifying and discriminating seismic events and interference signals based on Long Short-Term Memory (LSTM) networks. The method constructs the discrimination process as a supervised learning classification task, utilizing the temporal feature learning capability of LSTM to directly analyze the original three-component (or single-component) waveforms and output the event category (e.g., "earthquake" or "interference"). It employs an end-to-end intelligent analysis and discrimination technique using the LSTM network in a recurrent neural network to analyze and discriminate detected candidate seismic event waveforms. This intelligent seismic interference removal method automatically learns features from the original seismic waveforms, possesses temporal modeling capabilities, and improves classification accuracy.

[0075] Example 4: Based on the same inventive concept, the present invention also provides an interference signal removal system based on an earthquake interference signal removal model, comprising: Data acquisition module: When a seismic station issues a trigger event report, it acquires the current waveform data from the trigger event report; Interference signal removal module: Uses a pre-built seismic interference signal removal model to remove interference signals from the seismic waveform data to obtain the event category corresponding to the current waveform data; Earthquake early warning module; used to determine whether the triggered event report is a real earthquake event based on the event category, and if so, to trigger an earthquake early warning; The earthquake interference signal elimination model is constructed using an earthquake interference signal elimination model construction method described in Example 1.

[0076] Specifically, in this embodiment of the invention, the earthquake early warning module is used to: determine that the triggered event report is a real earthquake event when the trigger probability of the event category is greater than a preset threshold, and then trigger an earthquake early warning; and determine that the triggered event report is an interference event when the trigger probability of the event category is not greater than the preset threshold, and then not trigger an earthquake early warning.

[0077] Example 5: Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0078] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to implement the corresponding method flow or corresponding function, so as to implement the steps of the earthquake interference signal elimination model construction method in the above embodiments and the steps of the interference signal elimination method based on the earthquake interference signal elimination model in the above embodiments.

[0079] Example 6: Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device, used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the earthquake interference signal elimination model construction method in the above embodiments, as well as the steps of the interference signal elimination method based on the earthquake interference signal elimination model in the above embodiments.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for constructing a seismic interference signal removal model, characterized in that, include; Historical waveform data of triggering events are obtained from a historical earthquake triggering database; a training sample set is generated based on the annotation results of the historical waveform data, the training sample set including the historical waveform data and its corresponding annotation categories; The training sample set is input into a long short-term memory network, and the long short-term memory network is used to extract the temporal features of the historical waveform data to obtain the temporal features of the seismic signal. Based on the temporal characteristics of the earthquake signal, a feature-weighted aggregation algorithm is used to calculate the aggregated features of the earthquake signal; Based on the aggregation features of the seismic signals, a classification decision algorithm is used to calculate the predicted category of the historical waveform data. Based on the historical waveform data, the labeled categories, and the predicted categories, the long short-term memory network is iteratively trained to obtain a trained earthquake interference signal removal model.

2. The method according to claim 1, characterized in that, Historical waveform data of triggering events are obtained from the historical earthquake triggering database, including: Data on multiple triggered earthquake events were obtained from the historical earthquake triggering database of the historical seismic network. For each triggered earthquake event, the historical waveform data of the event is extracted using a fixed-length time window based on the triggering time. The historical waveform data includes vertical, north-south, and east-west components.

3. The method according to claim 1, characterized in that, The generation of the training sample set based on the annotation results of the historical waveform data includes: The historical waveform data is reviewed and labeled, and based on the review and labeling results, the corresponding earthquake event category label is labeled for each time window data to generate a training sample set; Specifically, when the earthquake event category label value is 1, it indicates that an earthquake event has occurred; When the earthquake event category label value is 0, it indicates that an interference event has occurred.

4. The method according to claim 1, characterized in that, Before inputting the training sample set into the Long Short-Term Memory network, the method further includes: The mean and linear detrending processes are performed on all components of each historical waveform data in the training sample set, and a bandpass filter is applied for noise reduction to obtain each historical waveform data after noise reduction. All components in each historical waveform data after noise reduction are normalized.

5. The method according to claim 1, characterized in that, The step of extracting seismic signal temporal features from historical waveform data using the long short-term memory network includes: A temporal feature extraction module employing a long short-term memory network model extracts the seismic signal temporal features of the historical waveform data in the time dimension; the seismic signal temporal features are used to identify seismic signals.

6. The method according to claim 1, characterized in that, The step of calculating the aggregated features of the seismic signal using a feature-weighted aggregation algorithm based on the time-series characteristics of the seismic signal includes: Based on the temporal characteristics of the earthquake signal, an attention mechanism algorithm is used to calculate the attention weight at each moment in the temporal characteristics of the earthquake signal. Based on the attention weights and the temporal features of the seismic signal at each time step, the feature weighting aggregation module of the long short-term memory network model is used to calculate the aggregated features of the seismic signal.

7. The method according to claim 6, characterized in that, The function expression for the feature weighted aggregation module is: in, Let be the attention weight at time t; Z represents the hidden state at time t; Z represents the seismic signal aggregation feature.

8. The method according to claim 1, characterized in that, The step of calculating the predicted category of the historical waveform data based on the aggregated features of the seismic signal using a classification decision algorithm includes: The aggregated seismic signal features are processed by nonlinear activation through the fully connected layer of the long short-term memory network model to obtain seismic feature spatial data. Based on the aforementioned seismic feature spatial data, the classification decision module of the Long Short-Term Memory Network model is used to determine the predicted category of the historical waveform data.

9. A system for constructing a seismic interference signal removal model, characterized in that, include: Data acquisition module: used to obtain historical waveform data of triggering events from the historical earthquake triggering database; Annotation module: used to generate a training sample set based on the annotation results of the historical waveform data, the training sample set including the historical waveform data and its corresponding annotation categories; Temporal feature extraction module: used to input the training sample set into the long short-term memory network, and use the long short-term memory network to extract temporal features from the historical waveform data to obtain the temporal features of the seismic signal; Feature aggregation module: used to calculate the aggregated features of the seismic signal based on the time-series features of the seismic signal using a feature-weighted aggregation algorithm; Classification and decision module: used to obtain the predicted category of the calculated historical waveform data based on the aggregated features of the seismic signal through a classification and decision algorithm; Training module: used to iteratively train the long short-term memory network based on the historical waveform data, the labeled categories, and the predicted categories to obtain a trained earthquake interference signal removal model.

10. The system according to claim 9, characterized in that, The data acquisition module is specifically used to acquire multiple triggered earthquake event record data from the historical earthquake triggering database of the historical earthquake network; For each triggered earthquake event, the historical waveform data of the event is extracted using a fixed-length time window based on the triggering time. The historical waveform data includes vertical, north-south, and east-west components.

11. The system according to claim 9, characterized in that, The annotation module is specifically used to review and annotate the historical waveform data, and based on the review results, to annotate the corresponding earthquake event category label for each time window data, thereby generating a training sample set. Specifically, when the earthquake event category label value is 1, it indicates that an earthquake event has occurred; When the earthquake event category label value is 0, it indicates that an interference event has occurred.

12. The system according to claim 9, characterized in that, The system also includes: The data preprocessing module is used to perform mean and linear detrending processing on all components of each historical waveform data in the training sample set, and to apply a bandpass filter for noise reduction. All components in each historical waveform data after noise reduction are normalized.

13. The system according to claim 9, characterized in that, The temporal feature extraction module is specifically used to extract the temporal features of the seismic signal in the time dimension using the long short-term memory network model temporal feature extraction module; the temporal features of the seismic signal are used to identify the seismic signal.

14. The system according to claim 9, characterized in that, The feature aggregation module is specifically used to calculate the attention weight at each moment in the seismic signal time series features based on the seismic signal time series features using an attention mechanism algorithm. Based on the attention weights and the temporal features of the seismic signal, a feature weighted aggregation module is used to calculate the aggregated features of the seismic signal.

15. The system according to claim 14, characterized in that, The function expression for the feature weighted aggregation module is: in, Let be the attention weight at time t; Z represents the hidden state at time t; Z represents the seismic signal aggregation feature.

16. The system according to claim 9, characterized in that, The classification decision module is specifically used to perform nonlinear activation processing on the aggregated features of the seismic signal through a fully connected layer to obtain seismic feature spatial data. Based on the aforementioned seismic feature spatial data, the predicted category of the historical waveform data is calculated using a classification decision module according to the classification type of the seismic signal.

17. A method for removing interference signals based on a seismic interference signal removal model, characterized in that, include; When a seismic station issues a trigger event report, the current waveform data is obtained from the trigger event report; The earthquake waveform data is processed by removing interference signals using a pre-built earthquake interference signal removal model to obtain the event category corresponding to the current waveform data. Based on the event category, determine whether the triggered event report is a real earthquake event; if so, trigger an earthquake early warning. The earthquake interference signal removal model is constructed using a method for constructing an earthquake interference signal removal model as described in any one of claims 1 to 8.

18. The method according to claim 17, characterized in that, The step of determining whether the triggered event report is a real earthquake event based on the event category includes: When the trigger probability of the event category is greater than a preset threshold, the trigger event report is determined to be a real earthquake event; When the trigger probability of the event category is not greater than a preset threshold, the trigger event report is determined to be an interference event.

19. An interference signal removal system based on a seismic interference signal removal model, characterized in that, include: Data acquisition module: When a seismic station issues a trigger event report, it acquires the current waveform data from the trigger event report; Interference signal removal module: Uses a pre-built seismic interference signal removal model to remove interference signals from the seismic waveform data to obtain the event category corresponding to the current waveform data; Earthquake early warning module; used to determine whether the triggered event report is a real earthquake event based on the event category, and if so, to trigger an earthquake early warning; The earthquake interference signal elimination model is constructed using a method for constructing an earthquake interference signal elimination model as described in any one of claims 1 to 8.

20. The system according to claim 19, characterized in that, The earthquake early warning module is specifically used to: determine that the triggered event report is a real earthquake event when the trigger probability of the event category is greater than a preset threshold, and then trigger an earthquake early warning; When the trigger probability of the event category is not greater than a preset threshold, the trigger event report is determined to be an interference event and no earthquake early warning is triggered.

21. A computer device, characterized in that, include: One or more processors; The processor is used to store one or more programs; When the one or more programs are executed by the one or more processors, they implement the earthquake interference signal removal model construction method as described in any one of claims 1 to 8 and the interference signal removal method based on the earthquake interference signal removal model as described in any one of claims 16 to 18.

22. A computer-readable storage medium, characterized in that, It contains a computer program, which, when executed, implements a method for constructing an earthquake interference signal removal model as described in any one of claims 1 to 8, and a method for removing interference signals based on an earthquake interference signal removal model as described in any one of claims 16 to 18.

Images