Ionized layer TEC pre-earthquake abnormal signal identification method based on artificial intelligence technology

Abstract

The invention relates to an ionosphere total electron content (TEC) pre-earthquake abnormal information extraction method based on an artificial intelligence technology. The method comprises the steps of S1, data preprocessing and sequence construction, S2, time sequence convolutional network (TCN) model construction with an auxiliary attention mechanism (AuxAttention), S3, model optimization and S4, earthquake monitoring application efficiency evaluation. According to the method, the TCN and the auxiliary attention mechanism are combined, so that the ionized layer TEC anomaly detection precision is greatly improved, the pre-earthquake anomaly signals can be effectively captured, and more accurate and timely early warning information is provided for earthquake prediction. The method shows higher prediction accuracy under different geographic positions and conditions, and is suitable for large-range earthquake monitoring application.

Description

Technical Field

[0001] The invention relates to the technical field of ionospheric TEC pre-earthquake abnormal signal recognition, and in particular to an ionospheric TEC pre-earthquake abnormal signal recognition method based on artificial intelligence technology. Background Art

[0002] With the development of artificial intelligence technology, ionospheric TEC prediction and abnormal signal recognition methods based on machine learning and deep learning have gradually become a research hotspot. In particular, the time series convolutional network (TCN), as an advanced deep learning technology, can effectively process and analyze time series data, and has potential advantages in the detection of ionospheric TEC anomalies. Through its powerful time series modeling capabilities, TCN can identify pre-earthquake TEC anomaly signals, thereby providing more accurate and timely earthquake warning information. This method based on artificial intelligence technology can not only improve the recognition accuracy of ionospheric abnormal signals, but also play an important role in the early warning of pre-earthquake abnormal signals. Summary of the invention

[0003] The purpose of the present invention is to provide an ionospheric TEC anomaly information extraction device based on a temporal convolutional network (TCN) and an auxiliary attention mechanism to solve the technical problem that the ionospheric TEC anomaly detection accuracy is not high in the prior art, especially the detection effect is poor in earthquake prediction applications.

[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions: including S1: data preprocessing and sequence construction, S2: construction of a temporal convolutional network (TCN) model with an auxiliary attention mechanism (AuxAttention), S3: model optimization, S4: earthquake monitoring application performance evaluation; Step S1: First, collecting historical data of ionospheric TEC is the basis of this method. In the past, TEC anomaly detection was mainly based on single stations, but due to the lack of density of single stations, it could not form a good detection of pre-earthquake TEC anomalies. Therefore, we improved it and used the temporal convolutional network (TCN) to detect regional TEC anomalies. First, we need to solve the observation values ​​of the stations in the region, and then establish the region. Among them, the estimation of code bias and receiver hardware delay is particularly important for the solution of TEC. When processing the differential code estimation of multi-system receivers, this study uses M_GIM software, which is an open source software based on M_DCB technology and is specially used for global and regional ionospheric modeling. M_GIM software is used to process the differential code estimation of multi-system receivers. The software is based on M_DCB technology and is designed for global and regional ionospheric modeling. Identify and process the C1C, C1C and C2W signal types in GPS pseudorange observation data, and concentrate on processing C1, C2 and P2 types of data. Before calculating the P4 smoothing value, code deviation correction is performed to ensure that the C1 data can accurately correspond to the P1 level and guarantee the accuracy and reliability of the analysis. Through this series of preprocessing steps, we can ensure the quality of the data and lay a solid foundation for the subsequent ionospheric TEC data analysis. After obtaining the dual-frequency observations of the observation station, the receiver coordinates are read, and the dual-frequency observation files are cycle-slip detected and repaired to obtain the smoothed P4 observations. The puncture point coordinates are calculated using the receiver coordinates and the satellite coordinates of the precise ephemeris. The normalized structural equation is constructed with the smoothed P4 observations. The DCB deviation file is used to estimate the DCB of the satellite and the receiver based on the least squares adjustment method to obtain the non-integral harmonic function model, the various parameters of the model, and the ionospheric regional model. As the modeling data set, the China regional TEC model data set is obtained. We process the data into a geographic grid with a time resolution of 2 hours and a spatial resolution of 1°×1° per grid, covering 15N°-55N°, 70E°-140E°, covering most of China, and combining it with spatial dimension information such as geographic location to form a complete time series. In addition to TEC data, related auxiliary data (KP, Dst, KP, F10.7) are also collected. These auxiliary data can provide additional information about the impact on the ionospheric state, which helps the model to more accurately identify TEC anomalies.

[0005] Step S2: After the data preprocessing in step S1, we build the model. The TCN module is the core of the model and is responsible for extracting the time series features of the ionospheric TEC data. TCN uses causal convolution and dilated convolution techniques to expand the convolution receptive field while ensuring the causality of the time series data, thereby capturing long-span dependencies. The specific implementation includes multiple layers of causal convolution. The activation function after each layer of convolution uses Leaky ReLU, which helps the model capture complex nonlinear relationships. At the same time, residual connections are used to avoid the gradient vanishing problem caused by deep networks. The features after each layer of convolution are filled with values ​​removed by the Chomp1d module to ensure true causality. In order to make full use of auxiliary data (such as geomagnetic index, solar activity index, etc.), the model introduces an auxiliary attention mechanism module. This module generates attention weights through two layers of fully connected networks, uses the Tanh activation function to capture nonlinear relationships, and normalizes the weights through the Softmax function to highlight key auxiliary features. The attention mechanism enhances the model's sensitivity to auxiliary information and improves the overall prediction performance. The TEC time series features extracted by TCN and the auxiliary features processed by the auxiliary attention mechanism are fused. The fused features are reduced in dimension through the fully connected layer, and the ReLU activation function is used to further extract nonlinear features. The Dropout layer is added to prevent overfitting, and finally the TEC outliers are predicted through the linear output layer. The loss function of the model uses the mean square error (MSE), and the optimization algorithm uses the Adam optimizer to ensure the stability of the model and the prediction accuracy.

[0006] Step S3: During the hyperparameter optimization process, Optuna was used to adjust the parameters of the TCN model, including kernel size, dropout rate, and number of TCN layers. By comparing different hyperparameter combinations, such as the number of TCN layers and auxiliary input dimensions, the model was optimized after each round of training using the AdamW optimizer and the CosineAnnealingWarmRestarts learning rate scheduler. After 40 training cycles, the mean squared error (MSE) of the model on the test set was evaluated as a validation metric, and Optuna was used to automatically search for the optimal hyperparameter combination to minimize the validation error and improve prediction performance.

[0007] S4: Evaluation of earthquake monitoring application performance. In the study of earthquake remote sensing monitoring, we first need to compare the predicted total electron content (TEC) with the measured TEC. The key to this process is to set a reasonable threshold to effectively extract the anomalies of ionospheric TEC. These anomalies may be associated with earthquake activity, so they provide important clues for earthquake monitoring. Next, we will carry out earthquake remote sensing monitoring applications for the extracted anomalies. In this step, we will select actual earthquake cases for verification to evaluate the accuracy and effectiveness of the model. By comparing the predicted results with the actual earthquake situation, we can get direct feedback on the performance of the model. Based on the feedback information from these application evaluations, we will continuously adjust the model parameters to optimize the model's prediction ability. This process is an iterative cycle: after each evaluation, we will adjust the model parameters according to the feedback results to gradually improve the accuracy of the model. In addition, we will use a training set to train the model and establish a basic prediction model. Subsequently, actual earthquake monitoring applications will be carried out and the effect will be continuously evaluated. The evaluation results after each application will provide us with valuable information to help us further optimize the model and enhance its application value in earthquake monitoring. Through this series of steps, we can not only conduct earthquake monitoring effectively, but also realize dynamic optimization of the model to adapt it to different monitoring needs and environmental changes, thereby improving the scientificity and practicality of the entire project.

[0008] Compared with the prior art, the TEC anomaly detection method proposed in the present invention significantly improves the detection accuracy of ionospheric TEC anomalies in earthquake prediction by combining TCN with an auxiliary attention mechanism. By combining a variety of auxiliary data (geomagnetic index and solar activity data), the model can more comprehensively capture the factors affecting TEC and improve the overall prediction performance. The model shows higher prediction accuracy in different geographical locations and conditions, can accurately capture ionospheric anomalies over a large range, and is suitable for a wide range of earthquake monitoring applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is the site distribution map; Figure 2 To model the flow chart; Figure 3 This is the architecture diagram of the TCN ionospheric TEC prediction model. DETAILED DESCRIPTION

[0010] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example

[0011] (1) Establishment of the China regional TEC model dataset: We use the Chinese mainland tectonic environment monitoring network. The Chinese mainland tectonic environment monitoring network (hereinafter referred to as the Land State Network) is a national-level comprehensive earth science observation network that uses the global satellite positioning system (GNSS) as the main technology, supplemented by space technologies such as very long baseline interferometry (VLBI), satellite laser ranging (SLR) and interferometric synthetic aperture radar (INSAR), combined with precise gravity and precise leveling observation technology to monitor the changes in the earth's lithosphere, hydrosphere and atmosphere in real time. The Land State Network is composed of a data system consisting of 260 continuously observed base stations, 2,000 irregularly observed regional stations, 1 national data center and 5 data sharing subsystems. It mainly serves earthquake prediction, and also serves military mapping, geodesy and meteorological forecasting, taking into account scientific research, social disaster reduction and national economic construction. This time, we used the observation data of 260 continuously observed base stations of the Land State Network for modeling.

[0012] Extraction Since the ionospheric delay depends on the frequency of the observable value, the first-order delay can be isolated by dual-frequency pseudorange or carrier phase measurement, while the high-order ionospheric refraction effect is usually ignored, which accounts for less than 0.1% of the ionospheric delay. After obtaining the dual-frequency observations of the observation station, the receiver coordinates are read, and the dual-frequency observation file is cycle slip detected and repaired to obtain the smoothed P4 observations. The puncture point coordinates are calculated using the receiver coordinates and the satellite coordinates of the precise ephemeris. The normalized structural equation is constructed with the smoothed P4 observations. The DCB deviation file is used to estimate the DCB of the satellite and the receiver based on the least squares adjustment method to obtain the non-integral harmonic function model, the various parameters of the model, and the ionospheric regional model. As the modeling data set, the TEC model data set of the Chinese region is obtained. Related auxiliary data (KP, Dst, KP, F10.7) are also collected. These auxiliary data can provide additional information about the impact on the ionospheric state, which helps the model to more accurately identify TEC anomalies.

[0013] (2) Establishment of the Temporal Convolutional Network (TCN) model with AuxAttention mechanism: The architecture of this model is a temporal convolutional network (TCN) with an auxiliary attention mechanism (AuxAttention). It mainly consists of the following parts: Temporal Convolutional Network (TCN): This part is the main feature extractor of the model, using multiple layers of one-dimensional convolutional layers (TemporalConvNet), where each layer of convolution has a different expansion coefficient to capture dependencies in different time ranges. Each layer of convolution is connected through residual connections to prevent gradient disappearance, and the Chomp1d class is used to crop the convolution results to ensure time series alignment.

[0014] Auxiliary Attention Mechanism (AuxAttention): It is used to process additional auxiliary data, which are weighted by the attention mechanism and combined with the output of TCN. This part of the module calculates the attention weights of the auxiliary data through the fully connected layer (Linear) and Softmax, and then applies these weights to the auxiliary data.

[0015] Data fusion and output: The outputs of the TCN and auxiliary attention modules are combined through concatenation and then processed through a fully connected layer (combine_linear). Finally, the fused features are output through the output_linear layer to output the model's prediction results.

[0016] Among them, TemporalConvNet: contains multiple temporal convolution blocks, each block is combined with Chomp1d, LeakyReLU and Dropout layers through two convolutional layers, and finally the input is directly added to the output through residual connection. AuxAttention: Use fully connected layers and Tanh, Softmax to generate attention weights, weight the auxiliary data, and combine it with the TCN output. Fusion layer: After combining the output of TCN with the auxiliary data, it is fused through a fully connected layer before making the final prediction. This architecture captures the temporal dependencies of the input sequence through TCN, and processes additional auxiliary information through the attention mechanism, combining the features of both to make the final prediction.

[0017] This architecture diagram shows how to incorporate auxiliary data in a time series model and dynamically select important auxiliary features through an attention mechanism to make the model more robust to multimodal data.

[0018] Input layer (Input): Main input data, shape is (batch_size, channels, seq_length), representing data of multiple time steps. Aux: Auxiliary data, shape is (batch_size, aux_size), representing additional information not directly related to the main input. Temporal ConvNet (TCN): TCN is composed of multiple TemporalBlocks, each of which is a module with stacked convolutional layers: Convolutional layer (Conv1d): One-dimensional convolution operation to extract time series features. Chomp1d: Cropping operation to ensure that the output size after convolution is aligned with the input. LeakyReLU: Non-linear activation function that gives the model non-linear modeling capabilities. Dropout: Randomly discard a part of neurons to prevent overfitting. Residual Connection: Add the input directly to the output of the convolutional layer to alleviate the gradient vanishing problem. Output of TCN: Take the output of the last time step, which represents the characteristics of the time series.

[0019] Auxiliary attention mechanism (AuxAttention): This module is dedicated to processing auxiliary data (aux). Its structure is as follows: Fully connected layer (Linear): Linear transformation of input auxiliary data. Tanh activation function: Nonlinear processing of linear transformation results. Softmax layer: Calculate the attention weight of auxiliary data, which is used to select the most relevant part of auxiliary information. The output of this module is weighted auxiliary features.

[0020] Concatenation: Concatenate the output of TCN and the output of the auxiliary attention mechanism to form a new feature vector. After concatenation, the feature dimension is num_channels[-1] * 2.

[0021] Feature Combination Layer: Fully connected layer (Linear): Linearly transform the concatenated features and compress them into a smaller feature space, so that they can better fuse data from two different sources. ReLU activation function: Nonlinearize the fused features.

[0022] Output Layer: Linear: Maps the fused features to the target space and outputs the final predicted value. The shape of the output is the same as the input data and is used for regression tasks to predict the future values ​​of time series. Loss Function: Use the mean square error loss (MSELoss) to measure the difference between the model's predicted value and the true value. Optimizer & Scheduler: AdamW: Adam optimizer with weight decay, used to adjust the model weights.

[0023] CosineAnnealingWarmRestarts: A learning rate scheduler that periodically adjusts the learning rate during training to improve convergence speed. Model training: The model is trained through multiple epoch iterations. In each epoch, the model generates predictions based on the input data, calculates the loss, and updates the weights through backpropagation. Gradient clipping is used to control the size of the gradient to prevent the gradient explosion problem.

[0024] (3) Hyperparameter optimization of the model: Use Optuna for hyperparameter optimization to minimize the validation loss of a deep learning model using a trial-and-error approach. Optuna is a state-of-the-art hyperparameter optimization framework that automates the process of searching for the best hyperparameter combination, especially for machine learning projects like this one. Optuna defines a search space for the model's hyperparameters, and then iteratively selects and evaluates different hyperparameter combinations based on the results of previous experiments. Here is a detailed description of the method and its benefits: Search space definition: The script defines the hyperparameters to be optimized, including: num_channels: Select from a set of predefined options such as (128, 256, 512), (128, 512, 1024), etc.

[0025] kernel_size: ranges from 2 to 5.

[0026] dropout: A floating point value between 0.1 and 0.5.

[0027] These hyperparameters control the complexity and capacity of the TCN (Temporal Convolutional Network) model as well as its dropout regularization.

[0028] Bayesian Search Algorithm: Optuna uses an intelligent search algorithm based on Bayesian optimization to balance exploration (trying new hyperparameter combinations) and exploitation (refining promising combinations). Specifically: Exploration enables the optimizer to sample various hyperparameters in the search space. Exploration refines parameters based on past results, focusing on configurations that have previously performed well. This allows Optuna to efficiently find the best configuration while reducing the total number of trials. Objective Function: The objective function in this script calculates the mean squared error (MSE) loss between the predicted output and the actual output. Optuna attempts to minimize this loss value. Early Stopping: Optuna can stop unsatisfactory trials early by automatically pruning them when performance is significantly worse, thereby reducing the waste of computational resources and time wasted on suboptimal configurations.

[0029] Advantages of this approach: Automation and flexibility: Optuna automates the hyperparameter tuning process, saving the time and effort required to manually select and test hyperparameters. Efficient exploration: By using Bayesian optimization, Optuna efficiently navigates the search space. This avoids the inefficiencies of methods such as grid search or random search, which either test all possible combinations or select randomly without learning from previous trials. Pruning poor trials: Optuna’s early stopping mechanism reduces computation time by terminating poor trials early, allowing resources to focus on promising configurations. Parallel execution: Optuna can easily parallelize trials, speeding up hyperparameter searches when using multiple CPUs or GPUs. Adaptive search: The search algorithm adapts over time, focusing more on the best performing hyperparameters, which helps improve model performance faster.

[0030] By optimizing hyperparameters such as the number of channels, convolution kernel size, and dropout rate, the model can achieve better performance while reducing validation loss, reducing overfitting and improving generalization ability.

[0031] (4) Evaluation of earthquake monitoring application effectiveness: To compare the predicted TEC with the measured TEC, we first need to compare the predicted total electron content (TEC) with the measured TEC. The key to this process is to set a reasonable threshold to effectively extract the anomalies of the ionospheric TEC. These anomalies may be associated with seismic activity, so they provide important clues for earthquake monitoring. Next, we will apply the extracted anomalies to earthquake remote sensing monitoring. In this step, we will select actual earthquake cases for verification to evaluate the accuracy and effectiveness of the model. By comparing the predicted results with the actual earthquake situation, we can get direct feedback on the performance of the model. Based on the feedback information from these application evaluations, we will continue to adjust the model parameters to optimize the model's prediction ability. This process is an iterative cycle: after each evaluation, we will adjust the model parameters based on the feedback results to gradually improve the accuracy of the model. In addition, we will use a training set to train the model and establish a basic prediction model. Subsequently, actual earthquake monitoring applications are carried out and the effect is continuously evaluated. The evaluation results after each application will provide us with valuable information to help us further optimize the model and enhance its application value in earthquake monitoring. Through this series of steps, we can not only conduct earthquake monitoring effectively, but also realize dynamic optimization of the model to adapt it to different monitoring needs and environmental changes, thereby improving the scientificity and practicality of the entire project.

[0032] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying abnormal signals before ionospheric TEC earthquakes based on artificial intelligence technology, characterized in that: Including S1: data preprocessing and sequence construction, S2: construction of a temporal convolutional network (TCN) model with auxiliary attention mechanism AuxAttention, S3: model optimization, S4: earthquake monitoring application performance evaluation; Step S1: First, collecting historical data of ionospheric TEC is the basis of this method. In the past, TEC anomaly detection was mainly based on single stations. However, due to the fact that the single stations were not densely deployed, they could not form a good detection of TEC anomalies before earthquakes. Therefore, we improved and used the time series convolutional network (TCN) to detect regional TEC anomalies. First, we need to solve the observation values ​​of the stations in the region, and then establish the region. Among them, the estimation of code bias and receiver hardware delay is particularly important for the solution of TEC. When processing the differential code estimation of multi-system receivers, this study used M_GIM software, which is an open source software based on M_DCB technology and is specially designed for global and regional ionospheric modeling. This software is based on M_DCB technology and is specially designed for global and regional ionospheric modeling. It identifies and processes C1C, C1C and C2W signal types in GPS pseudorange observation data, and centrally processes C1, C2 and P2 types of data. Before calculating the P4 smoothing value, code bias correction is performed to ensure that the C1 data can accurately correspond to the P1 level and ensure the accuracy and reliability of the analysis. Through this series of preprocessing steps, we can ensure the quality of the data. , laying a solid foundation for the subsequent ionospheric TEC data analysis; after obtaining the dual-frequency observations of the observation station, read the receiver coordinates, and perform cycle slip detection and repair on the dual-frequency observation files to obtain the smoothed P4 observations. The puncture point coordinates are calculated using the receiver coordinates and the satellite coordinates of the precise ephemeris, and the normalized structural equation is constructed with the smoothed P4 observations. The DCB deviation file is used to estimate the DCB of the satellite and the receiver based on the least squares adjustment method to obtain the non-integral harmonic function model, obtain the various parameters of the model, and obtain the ionospheric regional model as the modeling data. We collected data from the TEC model dataset for the Chinese region. We processed the data into a geographic grid with a time resolution of 2 hours and a spatial resolution of 1°×1° per grid, covering 15N°-55N°, 70E°-140E°, covering most of China. We combined the data with spatial dimension information such as geographic location to form a complete time series. In addition to TEC data, related auxiliary data KP, Dst, KP, and F10.7 were also collected. These auxiliary data can provide additional information about the impact on the ionosphere state, which helps the model to more accurately identify TEC anomalies. Step S2: After the data preprocessing in step S1, we build the model. The TCN module is the core of the model and is responsible for extracting the time series characteristics of the ionospheric TEC data. TCN uses causal convolution and dilated convolution techniques to expand the convolution receptive field while ensuring the causality of the time series data, thereby capturing the dependencies over a long time span. The specific implementation includes multiple layers of causal convolution, and the activation function after each layer of convolution uses Leaky ReLU helps the model capture complex nonlinear relationships. At the same time, residual connections are used to avoid the gradient vanishing problem caused by deep networks. The features after each convolution are filled with Chomp1d modules to ensure true causality. In order to make full use of auxiliary data such as geomagnetic index and solar activity index, the model introduces an auxiliary attention mechanism module, which generates attention weights through two layers of fully connected networks, uses Tanh activation function to capture nonlinear relationships, and normalizes weights through Softmax function to highlight key auxiliary features. The attention mechanism enhances the model's sensitivity to auxiliary information and improves the overall prediction performance. The TEC time series features extracted by TCN and the auxiliary features processed by the auxiliary attention mechanism are fused. The fused features are reduced in dimension through the fully connected layer, and the nonlinear features are further extracted using the ReLU activation function. The Dropout layer is added to prevent overfitting. Finally, the TEC outliers are predicted through the linear output layer. The loss function of the model uses the mean square error MSE, and the optimization algorithm uses the Adam optimizer to ensure the stability of the model and the prediction accuracy. Step S3: During the hyperparameter optimization process, Optuna was used to adjust the parameters of the TCN model, including kernel size, dropout rate, and number of TCN layers. By comparing different hyperparameter combinations, such as the number of TCN layers and auxiliary input dimensions, the model was optimized after each round of training using the AdamW optimizer and the CosineAnnealingWarmRestarts learning rate scheduler. After 40 training cycles, the mean square error (MSE) of the model on the test set was used as a validation metric for evaluation. Optuna was used to automatically search for the optimal hyperparameter combination, thereby minimizing the validation error and improving prediction performance. Step S4: Evaluation of earthquake monitoring application performance. In the study of earthquake remote sensing monitoring, we first need to compare the predicted total electron content TEC with the measured TEC. The key to this process is to set a reasonable threshold in order to effectively extract the anomalies of ionospheric TEC. These anomalies may be related to seismic activity. Therefore, they provide important clues for earthquake monitoring. Next, we will carry out earthquake remote sensing monitoring applications for the extracted anomalies. In this step, we will select actual earthquake cases for verification to evaluate the accuracy and effectiveness of the model. By comparing the predicted results with the actual earthquake situation, we can obtain direct feedback on the model performance. Based on the feedback information from these application evaluations, we will continue to adjust the model. We adjust the model parameters to optimize the model's predictive ability. This process is an iterative cycle: after each evaluation, we will adjust the model parameters based on the feedback results to gradually improve the model's accuracy; in addition, we will use a training set to train the model and establish a basic prediction model. Subsequently, we will carry out actual earthquake monitoring applications and continue to evaluate the effects. The evaluation results after each application will provide us with valuable information to help us further optimize the model and enhance its application value in earthquake monitoring. Through this series of steps, we can not only effectively carry out earthquake monitoring, but also realize dynamic optimization of the model to adapt it to different monitoring needs and environmental changes, thereby improving the scientificity and practicality of the entire project.

2. The method for identifying abnormal ionospheric TEC pre-earthquake signals based on artificial intelligence technology according to claim 1 is characterized in that: The temporal convolutional network (TCN) model with auxiliary attention mechanism AuxAttention mainly consists of a temporal convolutional network (TCN), an auxiliary attention mechanism AuxAttention, a data feature fusion layer and an output layer.

3. The method for identifying abnormal ionospheric TEC pre-earthquake signals based on artificial intelligence technology according to claim 2 is characterized in that: There are 2 TCN modules.

4. The method for identifying abnormal ionospheric TEC pre-earthquake signals based on artificial intelligence technology according to claim 2 is characterized in that: Concatenate the output of TCN and the output of the auxiliary attention mechanism to form a new feature vector; after concatenation, the feature dimension is num_channels[-1] * 2.

5. The method for identifying abnormal ionospheric TEC pre-earthquake signals based on artificial intelligence technology according to claim 2 is characterized in that: Convolution kernel kernel_size: is 3.

6. The method for identifying abnormal ionospheric TEC pre-earthquake signals based on artificial intelligence technology according to claim 2 is characterized in that: The China regional ionospheric model, which is processed by M_GIM software on the data of Lutai Network, is used as input data.

7. The method for identifying abnormal ionospheric TEC pre-earthquake signals based on artificial intelligence technology according to claim 2 is characterized in that: options = [ (128, 256, 512),].

8. The method for identifying abnormal ionospheric TEC pre-earthquake signals based on artificial intelligence technology according to claim 2 is characterized by: dropout = 0.3571606746344543.

9. The method for identifying abnormal ionospheric TEC pre-earthquake signals based on artificial intelligence technology according to claim 2 is characterized in that: aux_size = 4.

Images