A method and system for rapid detection of anomalous seismic signals

Abstract

The application discloses a kind of abnormal seismic electric signal fast detection method and system, system includes: training data making module, neural network building module, neural network training module, test application module;For geoelectric field time series data feature establishes abnormal seismic electric signal fast detection neural network model;Synthetic data is used to train neural network model;According to the defined network performance evaluation index, select appropriate training cycle;Real-time geoelectric data is input to the abnormal seismic electric signal fast detection neural network model that training is completed, to detect the effective abnormal seismic electric signal in geoelectric data that is submerged by noise, realize the fast and accurate detection of abnormal seismic electric signal that mass geoelectric field data is implied.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of seismic signal detection, and relates to a rapid detection technology of abnormal seismic electric signals based on LSTM-Former. BACKGROUND

[0002] The geoelectric field is an important geophysical field, and the abnormal signals related to the seismogenic process are contained in the continuous observation data of the geoelectric field, i.e. the seismic electric signals (SES). A large number of retrospective experimental evidences show that during the seismogenic process, the stress accumulation caused by plate movement will lead to micro changes in rock physical properties and geological structure, and the SES can carry and amplify these important micro information. At the same time, rock physics experiments also show that obvious SES can be recorded before the rock sample breaks. Therefore, it is of great importance to realize short-term (one year or less) earthquake prediction by using the time-frequency or amplitude characteristics of SES, so as to reduce earthquake casualties and economic losses.

[0003] However, it is difficult to detect and identify SES directly from the original data. On the one hand, the SES excited by the seismogenic process is usually weak, and the signal amplitude is much lower than the fluctuation amplitude of the geoelectric field noise. On the other hand, with the acceleration of industrialization, many cities have built rail transit systems and large power transmission equipment, and the artificial noise generated thereby causes a large amount of effective information in the data collected by the geoelectric station in a time period to be unable to be utilized. Therefore, how to suppress various mixed noises in the original data and extract effective abnormal signals has become a technical problem to be solved urgently.

[0004] Traditional seismic electric signal detection and recognition algorithms have been widely used in geoelectric field signal processing. Orihara et al. (2009) used independent component analysis algorithm to decompose geoelectric field data into multiple independent components to distinguish abnormal SES and rail transit noise, however, due to the fixed source signal separation design of this technology, its universality is poor. An Zhanghui et al. (2011) introduced Hilbert-Huang transform algorithm to suppress the interference of subway noise on geoelectric field data, but this method has shortcomings in preserving SES signal characteristics, affecting its recognition accuracy. Li Wei et al. (2013) and Liang Yue et al. (2015) both used wavelet analysis to decompose and reconstruct geoelectric field data for denoising, but the effect is highly dependent on the selection of key parameters of wavelet function and the setting of optimal decomposition level, lacking a unified standard. Varotsos et al. (2023) proposed to calculate the variance, entropy, normalized power spectrum and other features of geoelectric field data in natural time domain to distinguish SES and various noises, but the overall algorithm complexity is high, which is difficult to meet the efficient detection demand of SES in the big data era. Therefore, traditional algorithms have limitations such as dependence on fixed prior assumptions, expert experience, and slow detection speed.

[0005] Deep learning, as an end-to-end global optimization algorithm, has strong feature extraction and representation learning ability.

[0006] Existing work has introduced deep learning methods to suppress mixed noise and detect abnormal SES. However, compared with the wide application and significant achievements of deep learning in seismic phase picking, geophysical forward and inverse problems, its application in geoelectric field signal processing is still limited and has obvious defects. For example, Kanarachos et al. (2017) proposed an abnormal SES detection algorithm combining wavelet and fully connected network, but limited by the modeling ability of fully connected network, its input needs to be multi-level simplified feature data after wavelet decomposition, which limits the detection accuracy and flexibility of the algorithm. Wang Kaixiang et al. (2020) used long short-term memory network to suppress Gaussian noise, Poisson noise and noise of specific frequency in collected geoelectric data; but its noise simulation types are limited and single, which is quite different from the actual complex artificial noise, resulting in low detection accuracy of the algorithm in real data and insufficient practicality. Wei Lei (2021) used support vector machine to distinguish SES and geoelectric field noise, but limited by the small size of training set and the low proportion of positive samples, the algorithm performance is significantly limited.

[0007] In summary, existing abnormal SES detection technologies based on traditional and deep learning still have many limitations, and it is urgent to explore and build a more accurate and efficient SES detection technology scheme. SUMMARY

[0008] In order to overcome the above-mentioned deficiencies of the prior art, the present application provides an abnormal SES detection method and system based on an LSTM-Former autoencoder network, which integrates a long short-term memory (LSTM) module good at extracting time series data features and a multi-head attention mechanism of Transformer to construct an abnormal SES rapid detection neural network model; and uses the change characteristics of two typical SES (single transient SES and persistent SES) to construct a synthetic data set, so as to realize rapid and accurate detection of the abnormal SES hidden in the massive geoelectric field data.

[0009] The technical scheme adopted by the present application is:

[0010] An abnormal SES rapid detection algorithm based on an LSTM-Former autoencoder network, which establishes a matching neural network architecture (abnormal SES rapid detection neural network model) for geoelectric field time series data features; combines the typical change characteristics of SES to make synthetic data, and uses the synthetic data to train the neural network; selects a suitable training period according to the defined network performance evaluation index; inputs the measured geoelectric data into the trained abnormal SES rapid detection neural network model, so as to detect the effective abnormal SES hidden in the geoelectric data. The specific steps include:

[0011] A. Making a synthetic data set, including:

[0012] A1. Determine the geoelectric station to be processed, and select the time period without abnormal SES from the SES original data collected by the station;

[0013] A2. Divide the SES original data by day (24 hours), and eliminate invalid data with missing values;

[0014] A3. Filter the SES original data divided by day to obtain relatively "clean" SES background data;

[0015] A4. Combined with the change characteristics of two types of typical measured SES, randomly embed artificial simulated SES data in the same time of the SES original data and the "clean" SES background data;

[0016] Two types of measured SES change characteristics: 1. Single transient seismic electric signal (square wave with different amplitude and duration); 2. Persistent SES disturbance (multiple square waves superimposed).

[0017] A5. By adjusting the energy amplitude, duration and occurrence frequency of the simulated SES, repeatedly embed artificial simulated SES with different characteristics into the original and "clean" background data to construct a synthetic data set with diversity and representativeness.

[0018] The synthetic dataset samples include: containing N t Raw data of the geoelectric field at each sampling point x L (Network input data) and the corresponding "clean" background data y L (The labels corresponding to the network input data) Both have been embedded with simulated SES at the same time and location.

[0019] B. Establish an LSTM-Former autoencoder network (a neural network model for fast anomaly SES detection):

[0020] The anomaly SES fast detection neural network model constructed in this invention includes an LSTM module and a multi-head attention mechanism module. It is a seven-layer encoder-decoder structure with good signal compression and effective feature extraction capabilities, preserving key information while reducing dimensionality. The LSTM module is used to model the temporal dependencies in the geoelectric field data, which helps extract dynamic features in continuously changing processes. The multi-head attention mechanism module is used to capture long-distance dependencies between geoelectric field data at different time steps from multiple perspectives, and adaptively assigns higher attention weights to key time steps closely related to the task, thereby significantly improving the network model's modeling ability and prediction performance in anomaly SES detection tasks. The input and output data of this network model are the original geoelectric field data x embedded with simulated SES, respectively. L And the corresponding "clean" background field data y L The dimensions of the input and output layers are determined by the number of time sampling points N of the geoelectric field data. t Sure.

[0021] C. Select an appropriate training set period to train the network model:

[0022] Based on step A, a dataset is generated to train the network; the training data is fed into the network, and the network parameters are iteratively adjusted to reduce the error between the seismic electrical signal prediction results of the network model and the target model, until the network converges. Specifically, the synthesized network input data x... L Feeding this data into the LSTM-Former autoencoder network model yields the "clean" background field data predicted by the network. Subsequently, based on the prediction results and tag y L The loss function is calculated, and the network parameters are iteratively adjusted using the backpropagation algorithm to reduce the error of the loss function, so that the network prediction results are closer to the labels.

[0023] Two types of evaluation metrics are defined for network performance: root mean square percentage error (RMSE) and outlier detection success rate (SES). Based on the changes in these metrics during training, and considering both training time and overfitting, an appropriate training period is selected to complete network training.

[0024] The appropriate training period refers to: in the minimum training period (i.e. the shortest training time consumption), the root mean square percentage error index of the neural network model on the test set is the lowest, and the abnormal SES detection success rate index is the highest; at this time, the network parameters corresponding to the training period are selected.D. Network inversion effect evaluation:

[0025] According to the fitting error between the prediction results of the network on the test set data and the labels and the SES detection success rate, the network performance is quantitatively evaluated.

[0026] E. Measured abnormal SES detection;

[0027] For the geoelectric station measured data to be processed, first, pre-processing is performed, including removing invalid data and dividing the remaining data by day, then directly applying the network trained in step C to the station measured data to quickly and accurately detect abnormal SES.

[0028] According to the above steps, the geoelectric data anomaly detection based on the LSTM-Former self-encoder network can be completed, and the abnormal SES hidden in the strong noise in the geoelectric data can be quickly and accurately extracted.

[0029] As a preferred scheme, in step A1, in order to ensure that the selected original data does not contain abnormal SES excited by earthquakes as much as possible, according to the empirical relationship between magnitude M and the propagation range R of the seismic electromagnetic anomaly signal (Hattori, 2004):

[0030]

[0031] It is necessary to ensure that the selected target station data period is within 2 months before and after, so that the abnormal SES excited by earthquakes cannot propagate to the target station, and it is ensured that the selected original data does not contain abnormal SES excited by earthquakes.

[0032] As a preferred scheme, in step A2, if there are missing data in the daily divided data, the daily data does not participate in subsequent processing.

[0033] As a preferred scheme, in step A3, Savitzky-Golay filtering is used to process the original data, and it is stipulated that if the amplitude difference between two consecutive sample points in the signal is much smaller than the embedded simulated SES, it is considered that the data is already "clean" enough;

[0034] As a preferred embodiment, in step A4, referring to the characteristics of the measured SES, the two types of simulated SES include: (1) a single transient SES, which is simulated by a square wave, the amplitude range of which is set to 10-15 mV / km and the duration range is set to 25-30 min; (2) a continuous SES disturbance, which is composed of 5-10 square waves with shorter durations, the amplitude range of which is set to 10-15 mV / km and the duration range is set to 5-10 min.

[0035] As a preferred embodiment, in step A4, two types of simulated SES are randomly embedded between 0:00 and 24:00.

[0036] As a preferred embodiment, in step A5, the amplitude, duration, and other characteristics of the SES are adjusted within the variable range of the parameters specified in A3, and the SES with different characteristics are embedded into the original and "clean" background data to enrich the training data.

[0037] As a preferred embodiment, in step B, the first layer of the LSTM-Former autoencoder network adopts a one-dimensional convolutional neural network. Without changing the causal relationship of the time series, the input data is sparsified, reducing the data dimensionality and the number of network parameters to be trained, thereby improving the network training speed.

[0038] As a preferred embodiment, in step B, the LSTM-Former autoencoder network introduces a bidirectional LSTM layer after the convolutional layer. This layer can simultaneously capture forward (historical) and backward (future) temporal information in the input data, thereby introducing a more comprehensive temporal dependency when predicting the current time step. Considering that the bidirectional LSTM layer has effectively integrated the forward and backward temporal features, in order to maintain the lightweight nature of the network structure, each subsequent layer of the network uses a unidirectional LSTM layer to compress and reconstruct the data, so as to reduce computational complexity while ensuring modeling accuracy.

[0039] As a preferred embodiment, in step B, the LSTM-Former autoencoder network employs the multi-head attention mechanism, a core module of the Transformer. Considering that the effective information processed by the bidirectional LSTM layer is no longer transmitted unidirectionally, it is not possible to simply select the output of a certain time step as the final output of the entire layer. Therefore, a multi-head attention mechanism is used to weight the outputs of all time steps of the bidirectional LSTM layer. This mechanism can capture time step features from different levels and angles in parallel from multiple subspaces. Time step outputs that significantly contribute to completing the SES detection task will be adaptively assigned higher weights, thereby effectively improving the network's detection performance and generalization ability.

[0040] As a preferred solution, in step B, the LSTM-Former autoencoder network uses a fully connected layer to restore the compressed features to the same dimension as the input data.

[0041] As a preferred solution, in step C, the loss function used by the LSTM-Former autoencoder network training is:

[0042]

[0043] Where N represents the number of samples in the dataset, and is the amount of data relied on by the model training. Generally, the larger the data volume, the more comprehensive the data features contained, and the stronger the prediction performance of the trained network. represents the label of the i-th sample, i.e., embedding the simulated SES in the "clean" background data. represents the network prediction result of the i-th sample, which reflects the prediction ability of the network model on the sample. Ideally, the network prediction result should be infinitely close to the label. w represents the regularization coefficient, used to adjust the weight of the regularization term in the loss function to prevent model overfitting. ‖·‖1 represents the Manhattan distance, which is used to help produce sparse solutions in regularization. W represents the weight matrix to be trained in each layer of the network, which determines the connection strength between different neurons and is the core parameter of network model learning. b represents the bias vector to be trained in each layer of the network, which is used to adjust the output of the activation function to make the model have stronger expression ability.

[0044] As a preferred solution, in step C, the Adam optimization algorithm is used to implement network training, with an initial learning rate of 0.001 and a batch size of 128. During training, Dropout is used to randomly turn off some neurons to increase the generalization ability of the model. The entire training process is set to 200 cycles, and the test set is predicted at the end of each cycle to monitor the network training effect.

[0045] As a preferred solution, in step C, the Python programming language open source libraries Tensorflow and Kears are used to support all deep learning concepts and optimization algorithms, and the required central processing unit and graphics card are AMD Ryzen 73700X (16GB memory) and NVIDIA GeForce RTX 3070 (8GB video memory), respectively.

[0046] As a preferred solution, in step D, the root mean square percentage error (RMSPE) and the detection success rate of SES are used to evaluate the detection performance of the network, and the RMSPE is defined as follows:

[0047]

[0048] To quantitatively evaluate the detection success rate of extracted SES, the SES successfully detected needs to meet two conditions:

[0049]

[0050] Where ST and ET represent the start and end time of the network predicted SES, ST and ET represent the start and end time of the embedded simulated SES, and τ represents the duration of the simulated SES. N N S S S Equation (4) shows that if the above two defined time differences are both less than half of the duration of the simulated SES, it means that the SES is successfully extracted by the network. Therefore, the extraction success rate is defined as:

[0051]

[0052] Where M1 and M2 represent the number of SES successfully extracted by the network and the total number of SES, respectively.

[0053] In specific implementation, the application realizes an abnormal seismic electric signal rapid detection system based on LSTM-Former, which can accurately detect pre-seismic abnormal SES submerged by strong noise in real time, provide electromagnetic precursor information for short-term earthquake prediction, and realize intelligent analysis of geoelectric field data. The system includes a training data making module, a neural network building module, a neural network training module, and a test application module. The three modules are progressive, and the next module can be executed only after the previous module is completed. The training data making module is used to make the synthetic data set required for training the neural network model; the neural network building module is used to build the LSTM-Former network; the neural network training module reads in the synthetic data set and the built LSTM-Former network, optimizes the neural network according to the set loss function, makes the network prediction result infinitely close to the label, and then retains the optimal network parameters obtained by training; finally, the test application module applies the trained network to the test set synthetic data and measured data, calculates and retains the performance evaluation indicators of the network on the two types of data.

[0054] Compared with the prior art, the application has the following beneficial effects:

[0055] ​​​​The abnormal SES fast detection neural network model based on the LSTM-Former self-encoder network constructed by the method can accurately and quickly detect the abnormal SES before an earthquake from the geoelectric data disturbed by strong aliasing noise. The network does not need to rely on the premise assumption that the noise source is fixed, and even in the complex layout condition that the geoelectric station measuring line intersects with the train track, the network can also maintain high extraction accuracy, has good robustness, and effectively adapts to the measured data of the geoelectric station at any layout position; meanwhile, the detection efficiency of the abnormal SES in the geoelectric data is significantly improved, and the abnormal detection of 24 h collected data can be completed in tens of milliseconds on an ordinary personal computer; in addition, the network does not need manual intervention and parameter adjustment in the prediction stage, has high automation degree, is convenient to operate, is easy to integrate and deploy, and is conducive to the popularization and application of the algorithm. The experimental results show that the network can realize all-weather intelligent detection of the abnormal SES before an earthquake, and provides a reliable data-driven solution for intelligent analysis of geoelectric field data in earthquake prediction. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is a schematic diagram of two types (single transient SES (a) and persistent SES disturbance (b)) of synthetic data samples used for training the network by using the original data and "clean" background data in the specific implementation of the method. Among them, (a) and (b) respectively represent sample examples embedded with single transient SES and persistent SES disturbance. The four graphs from top to bottom in (a) or (b) respectively represent the original data, the "clean" data, the network input data embedded with the abnormal SES and the label.

[0057] Figure 2 is a structural schematic diagram of the abnormal SES fast detection multi-scale convolutional neural network model based on the LSTM-Former self-encoder constructed by the method.

[0058] Figure 3 is the trend of the RMSPE (black) and the SES extraction success rate (gray) of the network model constructed by the method in the training set (dotted line) and the test set (solid line) with the change of the training period for two types of training data samples (single transient SES (a) and persistent SES disturbance (b)). The black solid points in (a) and (b) respectively represent the 178th and 142nd training periods.

[0059] Figure 4 is an example of the detection result of the randomly extracted single transient SES sample in the test set by the detection network model constructed by the method. The three graphs from top to bottom in (a) or (b) respectively represent the embedded single transient SES, the network input data and the earthquake electric signal prediction result.

[0060] Figure 5are detection results of the detection network model constructed by the present application on randomly extracted persistent SES disturbance samples in the test set. DETAILED DESCRIPTION

[0061] The present application will be further described by examples in conjunction with the accompanying drawings, but the scope of the present application is not limited in any way.

[0062] The present application provides an abnormal seismic electric signal detection algorithm based on LSTM-Former, which comprises the following steps: 1) embedding two types of abnormal seismic electric signals (i.e. single transient seismic electric signal and persistent seismic electric signal disturbance) with typical change characteristics in the original geoelectric data containing aliasing noise and clean background geoelectric data respectively, and constructing a synthetic data set; 2) designing an LSTM-Former autoencoder network with symmetric topology integrating long short-term memory (LSTM) and Transformer multi-head attention mechanism according to the characteristics of the synthetic data set; 3) defining the root mean square percentage error and the detection success rate as performance evaluation indexes, determining the optimal training period by combining the change curves of the two indexes during the training process, and completing network training; 4) evaluating the network representation ability and detection effect based on the fitting degree of the test set prediction results and labels, and the detection success rate of abnormal seismic electric signals; 5) applying the trained network to the geoelectric station real-time data to realize rapid detection of abnormal electric signals related to the seismogenic process. The algorithm is efficient and robust, can suppress aliasing noise in one step, and supports all-weather and intelligent detection of abnormal seismic electric signals overwhelmed by noise in massive data.

[0063] The core of the abnormal seismic electric signal detection algorithm based on LSTM-Former proposed by the present application is as follows: 1. generating a synthetic sample set according to the abnormal SES change characteristics shown by the current retrospective research results and the noise characteristics of the original geoelectric field data; 2. building a matching LSTM-Former autoencoder network according to the characteristics of the synthetic data and the research target; 3. defining network performance evaluation indexes, considering the possible overfitting phenomenon of the network, weighing the performance expected to be achieved by the network and the calculation time, and selecting a suitable training period; 4. applying the trained network to the test set data first to evaluate its abnormal detection effect; finally, applying the network to the real-time data of the station.

[0064] The core technical contributions of the present application include the following three aspects: 1. The process method for making a synthetic data set. We make the training data set for training the LSTM-Former autoencoder network by the method described in step A. The data set made by this method overcomes the problems of low proportion and small scale of positive samples (i.e. abnormal SES) in the traditional abnormal SES detection data set, and the difference between the used noise and the actual complex artificial noise is large, and the two types of typical SES embedded basically cover the common SES variation characteristics. 2. Construction of LSTM-Former neural network model. First, the LSTM module is suitable for extracting the hidden features of the geoelectric field time series data; second, the autoencoder structure has excellent signal dimension compression and effective feature extraction capability; in addition, the multi-head attention mechanism can adaptively allocate weights, so that key features play a more important role in the task solving process, thereby significantly improving the prediction performance of the network. Therefore, we combine LSTM with multi-head attention mechanism to build a seven-layer LSTM-Former network with autoencoder architecture, which overcomes the problem of insufficient network modeling and representation ability in traditional abnormal SES detection. 3. Define effective evaluation indicators of network performance. We define an important indicator for evaluating the network model to detect abnormal SES, i.e. SES detection success rate. The standard for successful detection is to compare the time information of the network prediction with the artificially embedded abnormal SES (formula 4).

[0065] Through the above three original core technologies, the new algorithm solves the problems of excessive reliance on prior assumptions, low proportion of positive samples in training data, slow detection speed and low precision in traditional abnormal SES detection algorithm, and realizes accurate and efficient detection of abnormal SES.

[0066] A. Generate synthetic data set:

[0067] A1. Determine the location station to be processed, and select the time period without abnormal SES from the geoelectric field data collected at the station;

[0068] Two types of internationally recognized SES variation characteristics: 1. Single transient seismic electric signal Figure 1 a) - square wave with different amplitude and duration; 2. Persistent SES disturbance Figure 1 b) - multiple square waves superimposed. The selected station to be processed is a regional geoelectric field continuous observation station HKB. The data collected by this station has a default sampling frequency of 0.1 Hz (i.e. 8640 sampling point data in one day). The data used in this study includes the geoelectric field continuous observation data collected by HKB station from 1999 to 2000 Figure 1 (a) & (b): raw data).

[0069] A2. After selecting the data, divide the data of the two-year period by day, and eliminate invalid data with missing values;

[0070] A3. Use Savitzky-Golay filtering to clean the original data of the HKB station to obtain approximately "clean" background data Figure 1 (a) and (b): "clean" data);

[0071] Although the cleaned data still contains some noise, most of the artificial noise has been effectively suppressed, and abnormal SES can be clearly identified. Therefore, this study stipulates that if the amplitude difference between two consecutive sample points in the signal is much smaller than the embedded simulated SES, the data is considered to be sufficiently "clean".

[0072] A4. Two types of typical SES, single transient SES and persistent SES disturbance, are randomly embedded in the original data and "clean" data Figure 1 (a) and (b): input data and labels);

[0073] The amplitude range of single transient SES is set to 10-15 mV / km, and the duration range is set to 25-30 minutes; the amplitude range of persistent SES disturbance is set to 10-15 mV / km, and the duration range is set to 5-10 minutes, and the occurrence frequency is set to 5-10 times.

[0074] A5. Adjust the characteristics of SES within the set range, and repeatedly embed SES with different variation characteristics into the original and "clean" data, to enhance the diversity of training data and improve the generalization ability of the network.

[0075] B. Establish a seismic electric signal detection network model based on LSTM-Former:

[0076] The detection network model based on LSTM-Former constructed by the application adopts an encoder-decoder structure (as shown in Figure 2 The input layer node number of the LSTM-Former network is set to 8640 to match the daily input data dimension with a sampling rate of 0.1 Hz. After the input layer, the CNN layer is used for sparse processing of the data, which significantly reduces the data dimension and the number of network parameters to be trained while preserving the time sequence. Then, a bidirectional LSTM layer is introduced to fully utilize the information before and after the target time to constrain the output of the current time step. After the bidirectional LSTM layer, a multi-head attention mechanism is embedded to weight the output of all time steps of the bidirectional LSTM layer to enhance the attention to important time step information. Finally, the data compression and reconstruction are completed through a unidirectional LSTM layer and a fully connected layer.

[0077] C. Select a suitable training period to complete network training:

[0078] For this case, the number of training set samples and test set samples are set to 20000 and 2000, respectively. The loss function uses formula (2), and the Adam optimization algorithm is used to iteratively adjust the network parameters W and b in formula (2). At the same time, in order to alleviate the overfitting problem in the network training process as much as possible (i.e., the network over-learns the noise or details in the training data, and fails to effectively generalize the universal laws of the data, resulting in poor prediction effect of the network on the test set), random dropout is used to randomly turn off some neurons during training to increase the network generalization ability. The initial learning rate of the Adam algorithm is set to 1e-3, the learning decay rate is set to 1e-5, the batch size is set to 128, and the entire training process is set to 200 cycles.

[0079] The detection success rate of RMSPE and SES is used as an evaluation index to predict the test set after each training cycle to monitor the training effect of the network Figure 3 ); Considering the overfitting problem in the training process, the network parameters corresponding to the training cycle when the test set loss function is the smallest (178th and 142nd) are selected for subsequent analysis of the network inversion effect. At this time, the network training time corresponding to the 178th and 142nd cycle is 62.3 and 49.7 minutes, respectively.

[0080] D. Network inversion effect evaluation:

[0081] The trained network is applied to the test set to detect abnormal SES submerged by strong noise. The RMSPE values and detection success rates of abnormal SES of the 2000 samples in the test set are calculated. For single transient SES, they are 2.2% and 98.2%, respectively; for persistent SES disturbance, they are 4.7% and 99.0%, respectively. Two samples are randomly selected from the two test sets to show the network's detection results for abnormal SES Figure 4 &5).

[0082] Due to the strong interference of noise, the embedded simulation SES in the input test set data is difficult to observe directly. After inputting the test set data into the LSTM autoencoder, the position of the simulation SES can be clearly identified from the network prediction results. The accuracy of the detection results proves that the network has learned the mapping relationship between the input data and the label. The network does not need to manually adjust any parameters during the prediction stage, and it only takes 72ms to complete the abnormal SES detection of the data collected within 24h, which is sufficient to support the all-weather intelligent anomaly monitoring of the geoelectric field data.

[0083] The experimental results of the above cases prove that the detection network model based on the LSTM-Former self-encoder constructed by the application has strong nonlinear representation capability, can efficiently and accurately detect weak abnormal SES from massive geoelectric field data with strong noise interference, has generalization and robustness to SES with different change characteristics, the algorithm has high intelligence and low use threshold, and provides a reliable and automated solution for seismic precursor anomaly signal detection.

[0084] In summary, the core technology of the application includes: A: making a synthetic data set, this method of making a synthetic data set can 1. increase positive samples; 2. expand the training data set, 3. use measured noise data, which is more suitable for complex noise environment; 4. Two kinds of SES samples can increase the network generalization performance; B: build a neural network. 1. The input and output layers of the neural network must match the dimension of the training data; 2. The LSTM network layer inside the neural network is suitable for extracting the time sequence characteristics of the training data; 3. The use of attention mechanism inside the neural network can further improve the performance of the network in grabbing effective features. Therefore, we finally built a neural network as shown in Figure 2 C: Network performance evaluation method. Since the duration and occurrence time of SES are very important for earthquake prediction, we originally defined the SES detection success standard (formula 4); so as to realize the rapid detection of abnormal seismic electric signal based on LSTM-Former.

[0085] It should be noted that the purpose of publishing the embodiments is to help further understand the application, but those skilled in the art can understand that various replacements and modifications are possible without departing from the scope of the application and the appended claims. Therefore, the application should not be limited to the disclosed content of the embodiments, and the scope of protection claimed by the application is subject to the scope defined by the claims.

Claims

1. A method for rapid detection of anomalous seismic telecommunication signals, characterized in that, A neural network model for rapid detection of abnormal seismic electric signals (SES) is established according to the characteristics of the time series data of the geoelectric field; the neural network model is trained using the synthetic data; and the training period is selected according to the defined network performance evaluation index; The measured geoelectric data is input into the trained neural network model for rapid detection of abnormal seismic electric signals, so as to detect the effective abnormal seismic electric signals in the geoelectric data which are submerged by noise; The method comprises the following steps: A. Synthetic data set is prepared; comprising: In the same time of the SES original data and the clean SES background data, the simulated SES data is randomly embedded; By adjusting the energy amplitude, duration and occurrence frequency parameters of the simulated SES, different characteristics of the artificial simulated SES are repeatedly embedded into the original and clean background data to construct a synthetic data set; The synthetic data set sample comprises geoelectric field original data of a plurality of sampling points and corresponding clean background data; both have embedded simulated SES data at the same time position; B. An autoencoder network is established based on LSTM-Former, that is, a neural network model for rapid detection of abnormal SES: The established neural network model for rapid detection of abnormal SES comprises an LSTM module and a multi-head attention mechanism module; wherein the LSTM module is used to model the time series dependence in the geoelectric field data to extract dynamic features in the continuous change process; the multi-head attention mechanism module is used to capture the long-distance dependence between the geoelectric field data at different time steps from multiple angles, and to adaptively assign higher attention weights to the key time steps closely related to the task, thereby improving the modeling ability and prediction performance of the network model in the abnormal SES detection task; the input and output data of the network model are the geoelectric field original data embedded with simulated SES and the corresponding clean background field data; the dimension of the input and output layers is determined by the number of time sampling points of the geoelectric field data; C. Training the network model: The neural network model for rapid detection of abnormal SES is trained based on the data set generated in step A, and a trained neural network model is obtained; D. The inversion effect and performance of the neural network model for rapid detection of abnormal SES are quantitatively evaluated according to the fitting error between the prediction results of the network on the test set data and the labels and the SES detection success rate; E. The trained neural network model for rapid detection of abnormal SES is used for abnormal SES measurement, so as to rapidly and accurately extract the abnormal SES submerged by strong noise in the geoelectric data.

2. The rapid detection method for abnormal seismic electrical signals as described in claim 1, characterized in that, The SES original data is selected from the data period of 2 months before and after the target station, so that the original data selected does not contain the abnormal SES excited by the earthquake.

3. The rapid detection method for abnormal seismic electrical signals as described in claim 1, characterized in that, Specifically, the original data is processed using Savitzky-Golay filtering; and if the amplitude difference between two consecutive sample points in the signal is much smaller than the embedded simulated SES data, the data is defined as clean data.

4. The rapid detection method for abnormal seismic electrical signals as described in claim 1, characterized in that, In step A, the simulated SES data comprises single transient SES and persistent SES disturbance; the single transient SES is simulated by a square wave; and the persistent SES disturbance is superimposed by a plurality of square waves.

5. The rapid detection method for abnormal seismic electrical signals as described in claim 1, characterized in that, In step B, the constructed abnormal SES rapid detection neural network model is a seven-layer encoder-decoder structure; the first layer adopts a one-dimensional convolutional neural network, which performs sparse processing on the input data without changing the causal relationship of the time sequence, reduces the data dimension and the number of network parameters to be trained, and improves the network training speed; A bidirectional LSTM layer is introduced after the convolutional layer, which simultaneously captures the forward (historical) and backward (future) time information in the input data, thereby introducing comprehensive time sequence dependencies when predicting the current time step; each subsequent layer adopts a unidirectional LSTM layer to compress and reconstruct the data, reducing the computational complexity while ensuring modeling accuracy.

6. The rapid detection method for abnormal seismic electrical signals as described in claim 5, characterized in that, In step B, the abnormal SES rapid detection neural network model adopts the multi-head attention mechanism module of the Transformer to adaptively weight the outputs of all time steps of the bidirectional LSTM layer and capture time step features at different levels and angles from multiple subspaces in parallel.

7. The rapid detection method for abnormal seismic electrical signals as described in claim 6, characterized in that, The abnormal SES rapid detection neural network model uses a fully connected layer to restore the compressed features to the same dimension as the input data.

8. The rapid detection method for abnormal seismic electrical signals as described in claim 1, characterized in that, The loss function L(W, b) used by the abnormal SES rapid detection neural network model is represented as: where N is the number of samples of the data set; is the label of the i-th sample, i.e., the simulated SES data embedded in the clean background data; is the network prediction result of the i-th sample; λ w is the regularization coefficient; ‖·‖1 represents Manhattan distance; W represents the weight matrix to be trained in each layer of the network; and b represents the bias vector to be trained in each layer of the network.

9. The rapid detection method for abnormal seismic electrical signals as described in claim 1, characterized in that, In step D, the root mean square percentage error (RMSPE) and the detection success rate of SES are used to evaluate the detection performance of the network. Where the successfully detected SES satisfies two conditions, represented as: where ST N and ET N represent the start and end times of the network predicted SES, respectively, ST S and ET S represent the start and end times of the embedded simulated SES, respectively, and τ S represents the duration of the simulated SES. The success rate is defined as: Where M1 is the number of SES successfully extracted by the network model, and M2 is the total number of SES.

10. A system implemented by the method for rapid detection of abnormal seismic signals according to claim 1, characterized in that, The system includes a training data production module, a neural network building module, a neural network training module, and a test application module; wherein, The training data production module is used to produce synthetic data sets of seismic electric signals required for training the neural network model; The neural network building module is used to build an abnormal seismic electric signal rapid detection network model based on LSTM-Former; The neural network training module reads in the synthetic data set and the constructed abnormal seismic electric signal rapid detection network model, trains and optimizes the neural network according to the loss function, and obtains the optimal network parameters; The test application module is used to apply the trained network to the test set data and the measured data, and calculate the performance evaluation indicators of the network model.

Images