A pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events

Abstract

The present invention discloses a pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events. In order to fully consider the influence of earthquake cases on the time scale, a point processing algorithm is fused with a neural network, and earthquake case factors are fused with pre-earthquake anomaly physical quantities. First, a feature extraction layer for satellite TEC, satellite electromagnetic signals, and seismic events is proposed respectively, a fusion layer for the three, and an output layer. Finally, multiple earthquake cases are used for prediction, and the prediction results show that the combination of physical quantities and earthquake cases can predict the anomalies of seismic events with relatively high accuracy. Finally, it is shown that after feature fusion, the pre-earthquake anomaly ratio index increases by 2% - 5%.

Description

Technical Field

[0001] The invention belongs to the field of earthquake anomaly detection, and in particular relates to anomaly detection based on TEC and electromagnetic data. Background Art

[0002] China is located at the junction of the Eurasian Plate and the Indo-Australian Plate, and is also located on the Pacific Ring of Fire and the seismic belt, where the seismic activity belt is very active. In addition, China has complex geological structures, diverse topography, and very complex seismic structures and seismic activity. It is one of the most earthquake-prone areas in the world. The losses and impacts caused by earthquake disasters in China are also very huge. Research on earthquake precursors can help us to give early warnings and take measures to reduce the losses of earthquake disasters, thereby protecting people's lives and property. In addition, earthquake precursor research can also promote the development of related fields such as seismology and geophysics, and enhance our understanding of the internal structure of the earth and the mechanism of seismic activity. Geographical and geological studies have shown that a series of physical quantities will change before an earthquake occurs. And as time goes by, people can obtain more accurate and extensive data from monitoring these physical quantities, which is very important for the research on earthquake precursor anomaly detection. There are many physical quantities affected by earthquakes, among which the total electron concentration (TEC) of the ionosphere and electromagnetic signals have been shown to have a certain correlation with earthquakes in many studies. The present invention studies the detection of earthquake anomalies based on these two physical quantities. In addition, large earthquakes are often preceded and followed by accompanying earthquakes, so the impact of earthquake events on earthquake anomalies is also considered. Summary of the invention

[0003] In order to make full use of the characteristics of electromagnetic data and TEC data, the present invention will use a neural network to extract the characteristics of these two parts respectively. For the extraction of TEC features, Gaussian distribution is first used as a constant attention parameter to process the TEC original grid signal, and then a convolutional neural network and an LSTM neural network are combined to extract its features. For the extraction of electromagnetic features, a multi-layer LSTM network is used to extract its time series features. In addition, in order to consider the impact of earthquakes on data, a temporal point processing process (Temporal Point Process) is added to the network, and the parameters in the process are adjusted using the transfer learning method, and then the features are extracted through the fully connected layer. After extracting the features of the three different signals, a fully connected layer is used to connect the TEC and electromagnetic signal features, and the impact of event features is added on this basis. As a joint neural network training model, this framework resamples data for many years, normalizes the time axis, and then sends it to the framework for fusion.

[0004] In view of the above problems existing in the prior art, the present invention aims to provide a method for face image age conversion based on gradient adversarial attack and generative adversarial model. By using the gradient adversarial attack method to extract the prior knowledge of the decoder, discriminator, face recognition and age estimation modules, it is possible to directly generate seemingly real face age conversion pictures through iteration without additional training, and ensure that the face in the picture is as close as possible to the target age, and the expression, posture and background of the face remain unchanged.

[0005] To achieve the purpose of the present invention, the technical solutions adopted by the present invention are as follows:

[0006] A method for pre-earthquake anomaly detection based on the fusion analysis of TEC, electromagnetic data and seismic events, the method comprising the following steps:

[0007] Step 1: Construct a TEC neural network, an electromagnetic neural network, and a seismic case neural network respectively according to the data characteristics of TEC, electromagnetic data, and seismic cases;

[0008] Step 2: Construct the fusion layer and the output layer of the framework;

[0009] Step 3: For the pre-earthquake anomaly detection in a given area, first train the parameters of the seismic case feature extraction layer alone. After the parameter training in this area is completed, freeze this part of the parameters as the regional features of this area;

[0010] Step 4: Send the data processed in Step 1 into the network constructed in Step 2 to obtain the probability value of an earthquake occurring at the current moment.

[0011] In the said Step 1, the construction method of the TEC neural network is:

[0012] The input TEC spatial grid data has spatial features. Since anomalies often appear at the epicenter location, more attention needs to be placed on the central position of the picture in the subsequent network processing. Therefore, a fixed attention parameter is used for correction. The correction parameter used in this section is a two-dimensional Gaussian distribution curve.

[0013] After the function correction, the CNN network model will focus more attention on the anomalies at the epicenter location. Then, CNN (Convolutional Neural Network) is used to extract spatial features. CNN is a deep learning model specifically used for processing spatial data. When processing grid data, CNN can perform convolution operations through local area connections to extract local information features, and can make full use of the data features of spatial distribution. The input is X, the convolution kernel is K, and the output feature map after convolution is Y. Then the formula for the convolution operation is:

[0014]

[0015] Among them, i, j, and k represent pixel positions in the output feature map, n represents the size of the convolutional kernel, m represents the depth of the convolutional kernel, and ∈ is the bias term.

[0016] After weighted processing, for the construction of the TEC network, LSTM will be used to process the relevant information of its time series signals. Finally, the relevant features of the TEC network are extracted.

[0017] In step 1, the construction method of the electromagnetic neural network is as follows:

[0018] Since electromagnetic signals are time signals with a high sampling rate, this section mainly uses the Long Short-Term Memory (LSTM) neural network to extract their features. Compared with traditional RNNs, the LSTM network can better solve the problem of long-term dependencies, enabling the network to better understand and process long sequence data.

[0019] A multi-layer LSTM network can be implemented by connecting multiple LSTM units in series. Among them, the output of the previous layer serves as the input of the next layer, and the output of the last layer serves as the output of the entire network. A multi-layer LSTM network can better process complex sequence data, improving the expressive power and generalization ability of the model.

[0020] In step 1, the construction method of the earthquake case neural network is as follows:

[0021] The time series point processing method is a data processing method based on time series. It can identify key points in the time series and use these key points for data analysis and processing. In earthquake events, we can regard the occurrence of an earthquake as a key point and input other possible earthquakes before and after the earthquake as additional information into the model. In this way, the correlation between earthquake events can be better captured, improving the accuracy and reliability of earthquake impending detection.

[0022] The Hawkes point process can be used to model event sequences with self-exciting properties. Its basic idea is that the occurrence of each event may trigger the occurrence of subsequent events, and the occurrence of subsequent events may further trigger more events. This section will use the Hawkes point process to model the event sequence and predict the future event occurrence time and intensity through model learning. The model of the Hawkes point process can be expressed as:

[0023]

[0024] Where t is a time point on the time axis, λ(t) is the event intensity at time t, N t represents the number of events before t, μ is the base intensity, α is the influence intensity, h(t - t i ) represents from event ti The influence function between time \(t = 0\) and time \(t\). The influence function \(h(t)\) is usually represented by an exponential decay function:

[0025] h(t)=\omega e^{ -βt

[0026] where \(\omega\) is the amplitude of the influence function and \(\beta\) is the time constant of the influence function.

[0027] To learn the parameters of the Hawkes point process model, the maximum likelihood estimation method is used. Suppose the observed event sequence is \(S = t_1\), \(t_2\), …, \(t_n\), then the log-likelihood function of the Hawkes point process can be expressed as: 1 t_1 2 t_2 n t_n

[0028]

[0029] where \(\theta = (\mu,\alpha,\omega,\beta)\) are the model parameters and \(T\) is the upper bound of the observation time.

[0030] By maximizing the log-likelihood function, the stochastic gradient descent algorithm can be used to learn the model parameters.

[0031] In step 2, the fusion of electromagnetic data and TEC data is as follows:

[0032] Since the sampling frequency of electromagnetic data is one-dimensional time-series data per second, while the sampling frequency of TEC data is two-dimensional data per hour, due to their different input scales, it is necessary to transform their shapes at the output layer of their respective networks so that they can be concatenated at the concatenation layer. For the TEC network, when the output of the convolutional layer is obtained, the spatial image has been compressed into a one-dimensional time-series signal. Therefore, we use two fully connected layers to transform it. Through such a transformation, the electromagnetic data and TEC data can be connected at the concatenation layer to form a comprehensive neural network model.

[0033] In step 2, the fusion of earthquake event information is as follows:

[0034] Since the causes of earthquakes are complex and it is not certain that the impact of earthquakes on subsequent events is additive, more parameters are needed to ensure their impact on subsequent processes. Therefore, in order to make their changes superimposed in some unknown way and affect the electromagnetic and TEC data features, fully connected layers will be concatenated in the network part after point processing, and the scale range of the point processing signal will be transformed to the scale of the corresponding TEC and electromagnetic data.

[0035] In step 3, the learning process of the earthquake case network is as follows:

[0036] Since earthquake events rarely occur within a year, and the mechanisms of earthquake occurrence in different regions are different, when exploring the detection of multivariate information impending earthquake anomalies within a certain time range, the effect is often poor due to insufficient earthquake events. If earthquake events in multiple regions within a certain time range are selected, the detection effect will be poor due to different regional characteristics. Therefore, the method of transfer learning will be used to construct the parameters for earthquake event point processing. That is, decades of historical data in a certain research area will be used to set the parameters of the earthquake case network. After the adjustment is completed, these parameters will be frozen for future target training and will no longer change with training.

[0037] Therefore, in subsequent learning, only the parameters of the fully connected layer of the earthquake case network will be learned, which reduces the learning of parameters in the point processing process and reduces the learning time of the nonlinear part. It also reduces the data demand. For areas where only a few major earthquakes occur each year, the problem of insufficient data to fit the parameters of the point processing process is avoided. And because the parameter setting data is inconsistent with the training and testing data, the risk of overfitting is reduced to a certain extent, thereby improving the generalization of the model.

[0038] In step 4, the overall training process is:

[0039] TEC data, satellite electromagnetic data, and earthquake data are sent to the network in batches in groups of certain time periods. The calibration value of each group of data indicates that the 15 days before the earthquake are all abnormal time periods. The final output is compared with the calibrated data, and the overall learning method is gradient descent.

[0040] Compared with other methods, the method proposed in the present invention has the following beneficial effects:

[0041] First, the causes of earthquakes are relatively complex, and the present invention uses a method of using multiple data to complement each other for anomaly detection, namely TEC data, electromagnetic data, and earthquake case data, and makes full use of multiple data features to extract abnormal features of earthquake precursors.

[0042] Second, since earthquakes occur regionally, the earthquake case data used can be fixed to the earthquake cases in the selected area. This can better target the same area, fully taking into account the regionality of earthquakes, making the model more targeted.

[0043] Third, due to the small number of real earthquake cases, the transfer learning method is used for the learning of point processing parameters. The parameters are adjusted on a large number of earthquake case data, the data set is expanded, and these parameters are frozen when the final abnormal pre-earthquake precursor is judged. In addition, the training parameters can be reduced in the middle of the training process, speeding up the training process.

[0044] Fourth, the point processing process is incorporated into the feature extraction part, which improves the interpretability of the model. And since the feature extraction work is carried out separately, the impacts of the three data sources on the model can be observed separately, and separate detection results can be given. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is the overall structural diagram of the model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0047] Embodiment 1: Refer to Figure 1 , a pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events, characterized in that the method comprises the following steps:

[0048] Step 1: Construct a TEC neural network, an electromagnetic neural network, and a seismic case neural network respectively according to the data characteristics of TEC, electromagnetic data, and seismic cases;

[0049] Step 2: Construct the fusion layer and the output layer of the framework;

[0050] Step 3: For the pre-earthquake anomaly detection in a given area, first train the parameters of the seismic case feature extraction layer alone. After the parameter training in this area is completed, freeze this part of the parameters as the regional characteristics of this area;

[0051] Step 4: Send the data processed in Step 1 into the network constructed in Step 2 to obtain the probability value of an earthquake occurring at the current moment.

[0052] In the said Step 1, the TEC neural network is:

[0053] Since the TEC data has spatial characteristics and there is a certain correlation between the spatial characteristics and earthquakes, a Gaussian distribution function is used to correct the spatial TEC data. After the function correction, the subsequent spatial feature extraction module will focus more attention on the anomalies at the epicenter position.

[0054] In the said Step 1, the construction method of the TEC neural network is:

[0055] The input TEC spatial grid data has spatial characteristics. Since anomalies often appear at the epicenter position, more attention needs to be placed on the center position of the picture in the subsequent network processing. Therefore, a fixed attention parameter is used for correction. The correction parameter used in this section is a two-dimensional Gaussian distribution curve.

[0056] After the function is corrected, the CNN network model will focus more on the anomalies at the epicenter location. Then, CNN (Convolutional Neural Network) is used to extract spatial features. CNN is a deep learning model specifically designed for processing spatial data. When dealing with grid data, CNN can perform convolution operations through local area connections to extract local information features and make full use of the data features of spatial distribution. The formula for the convolution operation is:

[0057]

[0058] Among them, the input is X, the convolution kernel is K, the output feature map after convolution is Y, i, j, and k represent the pixel positions in the output feature map, n represents the size of the convolution kernel, m represents the depth of the convolution kernel, and ∈ is the bias term.

[0059] After weighted processing, for the construction of the TEC network, LSTM will be used to process the relevant information of its time series signal, and finally the relevant features of the TEC network will be extracted.

[0060] In step 1, the electromagnetic neural network is:

[0061] The electromagnetic neural network mainly has time series features. Since its sampling frequency is much higher than that of the TEC signal, a multi-layer LSTM network is used to extract these features.

[0062] In step 1, the construction method of the electromagnetic neural network is:

[0063] Since the electromagnetic signal is a time signal with a high sampling rate, in this section, the LSTM (Long Short-Term Memory) neural network is mainly used to extract its features. Compared with the traditional RNN, the LSTM network can better solve the problem of long-term dependence, enabling the network to better understand and process long sequence data.

[0064] The multi-layer LSTM network can be realized by connecting multiple LSTM units in series. Among them, the output of the previous layer is used as the input of the next layer, and the output of the last layer is used as the output of the entire network. The multi-layer LSTM network can better process complex sequence data and improve the expression ability and generalization ability of the model.

[0065] In step 1, the earthquake event neural network is:

[0066] Since the time series of earthquake events is different from that of the other two signals and is discontinuous, the Hawkes point processing method is incorporated into this network to fit earthquake events.

[0067] In step 1, the construction method of the earthquake case neural network is:

[0068] The time series point processing method is a data processing method based on time series. It can identify key points in the time series and use these key points for data analysis and processing. In seismic events, we can regard the occurrence of an earthquake as a key point and input other possible earthquakes before and after the earthquake as additional information into the model. In this way, the correlation between seismic events can be better captured, and the accuracy and reliability of earthquake impending detection can be improved.

[0069] The Hawkes point process can be used to model event sequences with self-exciting properties. The basic idea is that the occurrence of each event may trigger the occurrence of subsequent events, and the occurrence of subsequent events may further trigger more events. In this section, the Hawkes point process will be used to model the event sequence, and the future event occurrence time and intensity will be predicted through model learning. The model of the Hawkes point process can be expressed as:

[0070]

[0071] where t is a time point on the time axis, λ(t) is the event intensity at time t, N t represents the number of events before t, μ is the base intensity, α is the influence intensity, and h(t - t i ) represents the influence function from event t i to time t. The influence function h(t) is usually represented by an exponential decay function:

[0072] h(t) = we -βt

[0073] where ω is the amplitude of the influence function and β is the time constant of the influence function.

[0074] To learn the parameters of the Hawkes point process model, the maximum likelihood estimation method is used. Assuming that the observed event sequence is S = t 1 , t 2 , …, t n , then the log-likelihood function of the Hawkes point process can be expressed as:

[0075]

[0076] where θ = (μ, α, ω, β) are the model parameters and T is the upper bound of the observation time.

[0077] By maximizing the log-likelihood function, the stochastic gradient descent algorithm can be used to learn the model parameters.

[0078] TEC spatial feature extraction module

[0079] Since the TEC is grid data, a multi-layer CNN (Convolutional Neural Network) is used to extract spatial features. The CNN at the end reduces the dimension of the TEC signal to one dimension and uses it as the input for the next time-series feature extraction module.

[0080] Time-series feature extraction module

[0081] The TEC has obvious time-series features. Therefore, this module uses an LSTM network to extract these features and concatenates a fully connected layer after extraction as the output of the TEC neural network.

[0082] In step 2, the fusion layer and output layer of the framework are as follows:

[0083] For the fusion of the TEC and electromagnetic signals, a fully connected layer is first used to unify their dimensions, and then the data of the two are fused by concatenation.

[0084] For the fusion of its fusion result with the earthquake event, since the earthquake event uses transfer learning, a fully connected layer is needed to characterize the network and change its scale to be the same as that of the TEC and electromagnetic signals. The influence on the TEC and electromagnetic signals is fused by accumulation, and the final output uses a fully connected method to reduce the output dimension to one dimension and uses the softmax activation function to adjust the output.

[0085] In step 3, the learning process of the earthquake case network is as follows:

[0086] For earthquake events in a certain area, since the earthquake case dataset is small, but the time span of the earthquake case dataset is large. Compared with electromagnetic data and TEC data, earthquake case data can be traced back to 1900. Using all known earthquake case data in this area for training can characterize the regional characteristics of this area. After the module is trained, it is saved and frozen in subsequent training processes.

[0087] In step 4, the overall training process is as follows:

[0088] The TEC data, satellite electromagnetic data, and earthquake case data are fed into the network in batches at a certain time interval. The calibration value of each group of data indicates that the 15 days before the earthquake are abnormal time periods. The final output is compared with the calibrated data, and the overall learning is carried out using the gradient descent method.

[0089] Embodiment 2:

[0090] See Figure 1 , a pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and earthquake events, the method comprising the following steps:

[0091] Step 1: Construct a TEC neural network, an electromagnetic neural network, and a seismic event neural network respectively according to the data characteristics of TEC, electromagnetic data, and seismic event examples;

[0092] Step 2: Construct the fusion layer and the output layer of this framework;

[0093] Step 3: For the pre-seismic anomaly detection of earthquakes in a given area, first train the parameters of the seismic event feature extraction layer separately. After the parameter training in this area is completed, freeze these parameters as the regional features of this area;

[0094] Step 4: Send the data processed in Step 1 into the network constructed in Step 2 to obtain the probability value of an earthquake occurring at the current moment.

[0095] In Step 1, the TEC neural network model is:

[0096] The input is processed into 7×7 spatial grid data. The sampling rate is 1 hour. The TEC network model is designed as a three-layer convolutional layer with an input of 720 and an output of 720. The convolution kernel size is 3×3, and no convolution padding is performed. An RELU layer is added as the activation function after each convolution. After three layers of convolution, the original TEC data becomes 1×720 in size. Therefore, an LSTM network with an input of 720 and an output of 512 is required to extract this feature. Finally, a fully connected layer of 512×512 is used to reduce the data feature to 512 dimensions.

[0097] In Step 1, the electromagnetic neural network model is:

[0098] The sampling rate of the electromagnetic data is 1-second one-dimensional time series data. A five-layer LSTM neural network with an output of 1024 is used to extract this feature and reduce its feature dimension to 1024.

[0099] In Step 1, the seismic event network model is:

[0100] Seismic events are grouped by 30 days, and a total of 720 seismic events are extracted in hours. The point where the earthquake occurs is marked as 1. The seismic events use the Hawkes point process. Its three parameters μ, α, and β are necessary parameters for training. To avoid bad data during training, the softplus function is used to correct this parameter, and the formula is:

[0101] softplus(x) = ln(1 + e x )

[0102] In Step 2, the fusion of TEC and electromagnetic data is:

[0103] Since the sampling frequency of electromagnetic data is one-dimensional time-series data per second, while the sampling frequency of TEC data is two-dimensional spatial data per hour, due to their different input scales, it is necessary to transform their shapes in the respective network output layers so that they can be concatenated in the concatenation layer. For the TEC network, when the output of the convolutional layer is output, the 7×7 spatial image has already been compressed into a one-dimensional time-series signal. Therefore, two fully connected layers are used to transform it, where the first layer has 512 neurons and the second layer has 256 neurons. Through such a transformation, electromagnetic data and TEC data can be connected in the concatenation layer to form a comprehensive neural network model.

[0104] To concatenate electromagnetic data and TEC data, it is necessary to unify their output shapes. Since the sampling frequency of electromagnetic data is relatively high, four fully connected layers can be used to transform and fit its features, with sizes of 1024×1024, 1024×1024, 1024×512, and 1024×512 respectively, and its output shape is 512. After being processed by the concatenation layer, its feature dimension is compressed to 512. After concatenation processing, the data size is (1, 1024).

[0105] In step 2, the earthquake event fusion is as follows:

[0106] Two fully connected layers are used with parameters of 720×512 and 512×512 respectively. Its output is set to 512 and fused with TEC and electromagnetic signals in a superimposed manner. The final fusion output uses multiple fully connected layers with parameters expressed as 1024×512, 512×256, 256×128, 128×64, and 64×1. And finally, the softmax function is used to limit the output range between 0 and 1.

[0107] In step 3, the earthquake event learning process is as follows:

[0108] All earthquake case data in a certain area since 1900 are used to train the processing process at this point. After training, the training parameters are frozen. These three parameters will represent the earthquake activity index of this area and have a certain regionality. They are used in subsequent network training.

[0109] In step 4, the overall training process is as follows:

[0110] For the labels, all data within 15 days before the earthquake are regarded as abnormal data. And all data are divided into a training set and a test set according to a ratio of 7:3. The Adam algorithm is used for backpropagation. The initial learning rate is 0.1 and it decreases in a cosine decay manner.

[0111] It should be noted that the above embodiments are only embodiments of the present invention and are not used to limit the protection scope of the present invention. Any equivalent replacement or substitution made on the basis of the above technical solutions falls within the protection scope of the present invention.

Claims

1. A pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events, characterized in that, the method comprises the following steps: Step 1: Construct a TEC neural network, an electromagnetic neural network, and a seismic case neural network respectively according to the data characteristics of TEC, electromagnetic data, and seismic cases; Step 2: Construct a fusion layer and an output layer of the neural network; Step 3: For the pre-earthquake anomaly detection in a given area, first train the parameters of the seismic case feature extraction layer alone. After the parameter training in this area is completed, freeze this part of the parameters as the regional characteristics of this area; Step 4: Send the data processed in Step 1 into the network constructed in Step 2 to obtain the probability value of an earthquake occurring at the current moment; In Step 2, the fusion layer and the output layer of the framework are: For the fusion of TEC and electromagnetic signals, a fully connected layer is used to unify the dimensions of the two first, and then the data of the two are fused by using the splicing method. For the fusion of its fusion result and seismic events, since the seismic events adopt the method of transfer learning, a fully connected layer is needed to characterize the network, and its scale is changed to the same scale as that of TEC and electromagnetic signals. The influence on TEC and electromagnetic signals is fused by using the accumulation method. The final output uses the fully connected method to reduce the output dimension to one dimension and uses the softmax activation function to adjust the output.

2. The pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events according to claim 1, characterized in that: In Step 1, the construction method of the TEC neural network is: The input TEC spatial grid data has spatial characteristics. Since anomalies often appear at the epicenter location, more attention needs to be placed on the center position of the picture in the subsequent network processing. Therefore, a two-dimensional Gaussian distribution curve is used for correction. After correction, the downstream neurons will pay more attention to the grid center position. Then, a convolutional layer is used to extract spatial characteristics. The convolutional layer can perform convolutional operations through local area connections, so as to extract local information characteristics and can make full use of the data characteristics of spatial distribution. The formula for the convolutional operation is: Where the input is X, the convolutional kernel is K, the output feature map after convolution is Y, i, j, and k represent the pixel positions in the output feature map, n represents the size of the convolutional kernel, m represents the depth of the convolutional kernel, and ∈ is the bias term. After weighted processing, the construction of the TEC network will use LSTM to process the relevant information of its time series signal, and finally extract the relevant characteristics of the TEC network.

3. The pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events according to claim 1, characterized in that: In Step 1, the construction method of the electromagnetic neural network is: Since the electromagnetic signal is a time signal with a high sampling rate, an LSTM (Long Short-Term Memory) neural network is mainly used to extract its features. The multi-layer LSTM network is implemented by connecting multiple LSTM units in series. Among them, the output of the previous layer is used as the input of the next layer, and the output of the last layer is used as the output of the entire network. The multi-layer LSTM network can better process complex sequence data and improve the expression ability and generalization ability of the model.

4. A pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events according to claim 1, characterized in that: In step 1, the construction method of the seismic case neural network is: The Hawkes point process is used to model event sequences with self-exciting properties. Its basic idea is that the occurrence of each event may trigger the occurrence of subsequent events, and the occurrence of subsequent events may further trigger more events. In this section, the Hawkes point process will be used to model the event sequence, and the model learning will be used to predict the future event occurrence time and intensity. The model representation of the Hawkes point process is: where t is a time point on the time axis, λ(t) is the event intensity at time t, N t represents the number of events before t, μ is the base intensity, α is the impact intensity, h(t - t i ) represents the impact function from event t i to time t. The impact function h(t) is usually represented by an exponential decay function: h(t) = ωe -βt where ω is the amplitude of the influence function, and β is the time constant of the influence function, To learn the parameters of the Hawkes point process model, the maximum likelihood estimation method is used. Assuming that the observed event sequence is S = t 1 , t 2 ,..., t n , the log-likelihood function of the Hawkes point process is expressed as: where θ = (μ, α, ω, β) are the model parameters, and T is the upper bound of the observation time, The model parameters are learned using the stochastic gradient descent algorithm by maximizing the log-likelihood function.

5. A pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events according to claim 2, characterized in that: TEC spatial feature extraction module, Since TEC is grid data, a multi-layer CNN (convolutional neural network) is used to extract spatial features. The CNN at the end reduces the TEC signal dimension to one dimension and uses it as the input of the next temporal feature extraction module.

6. A pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events according to claim 3, characterized in that: Temporal feature extraction module, TEC has obvious temporal features, so this module uses an LSTM network to extract these features and concatenates a fully connected layer after extraction as the output of the TEC neural network.

7. A pre-earthquake anomaly detection method based on the fusion analysis of TEC, electromagnetic data, and seismic events according to claim 1, characterized in that: In step 4, the overall training process is: The TEC data, satellite electromagnetic data, and seismic case data are batch-fed into the network in a certain time period. The calibration value of each group of data indicates that the 15 days before the earthquake are abnormal time periods. The final output is compared with the calibrated data, and the overall learning is carried out using the gradient descent method.

Images