Low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network

Abstract

The invention discloses a low signal-to-noise ratio seismic detection model based on a deep residual shrinkage network, which introduces an attention mechanism on the basis of designing a soft threshold function according to different noises in seismic records collected by different stations or detectors, and constructs the deep residual shrinkage network. The method is used for detecting seismic events in low-signal-to-noise-ratio seismic records. In the designed network, a soft threshold function is inserted into a deep network structure as a nonlinear transformation layer so as to effectively filter the influence of noise-related features. Meanwhile, an attention mechanism is introduced into the network, a self-adaptive slope module is added, the optimal threshold value of seismic signal denoising is automatically obtained, and optimal processing of different signals and different degrees of noise is achieved. Through training, the deep residual shrinkage network can adaptively determine the de-noising threshold, so that each seismic signal has its own threshold set, the purpose of accurately detecting seismic events under strong background noise is achieved, and the accuracy of seismic event detection is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of earthquake detection, and in particular relates to a low signal-to-noise ratio earthquake detection model based on a deep residual shrinkage network. Background Art

[0002] Earthquake event detection refers to the identification of earthquake events from continuous earthquake record data, which is the core issue in seismological observation. Aftershock monitoring and rapid positioning after a major earthquake, tomographic imaging of the Earth's internal velocity structure, and the compilation of earthquake catalogs are all based on earthquake event detection. With the continuous expansion of the scale of seismic networks and mobile arrays and the exponential growth of density, earthquake observation data has exploded. Relying on human-computer interaction to detect earthquake events from massive earthquake records can no longer meet the requirements of intelligent processing of earthquake data in the big data era. Therefore, how to accurately and automatically detect earthquake events from earthquake records has become an urgent problem to be solved in the current processing and application of earthquake data.

[0003] In the prior art, deep learning methods have achieved certain results in earthquake event detection, but they have high requirements on the quality of earthquake data. For earthquake records with high signal-to-noise ratio, generally good detection results can be obtained, but for earthquake records with low signal-to-noise ratio (such as microseismic events, whose energy is extremely weak and may even be noise-level), the reliability of the detection results may be problematic. Due to the complexity and variability of the earthquake observation environment, the collected earthquake records are mixed with various noises, and the presence of these noises will seriously affect the accuracy of earthquake event detection.

[0004] Therefore, before detection, seismic records are usually denoised. Common seismic signal denoising methods include empirical mode decomposition and wavelet threshold denoising. The main principle of these methods is to transform the original signal to an unimportant near-zero domain, and then apply a threshold function to convert features close to zero to zero to achieve the purpose of denoising. Usually the denoising threshold of these methods is fixed. However, studies have found that the noise of seismic records collected by different stations or detectors varies greatly, which leads to poor denoising effects of such methods. Especially for seismic records with low signal-to-noise ratio, the denoising effect is even worse. For example, for microseismic signals generated by hydraulic fracturing, due to the complex monitoring environment and strong background noise interference, microseismic events with weak energy are submerged in the background noise. Therefore, for the detection of seismic events, especially the detection of microseismic events, inappropriate denoising will lead to lower detection effects.

[0005] The design of the threshold function is a key step in signal denoising. In order to improve the denoising effect, the soft threshold function has been introduced into many seismic signal denoising methods. However, setting a suitable soft threshold function has always been a very challenging task, requiring a lot of expertise in seismic signal processing. In addition, the noise of seismic records collected by different stations or detectors varies greatly, and the noise intensity is different, resulting in different optimal threshold functions for different signals, which makes the design of the soft threshold function more difficult. Summary of the invention

[0006] The purpose of the present invention is to provide a low signal-to-noise ratio earthquake detection model based on a deep residual shrinkage network. On the basis of designing a soft threshold function, an attention mechanism is introduced to construct a deep residual shrinkage network for detecting earthquake events in low signal-to-noise ratio earthquake records, which effectively improves the accuracy of earthquake event detection. In order to achieve the above purpose, the technical solution adopted by the present invention is specifically as follows:

[0007] A low signal-to-noise ratio earthquake detection model based on a deep residual shrinkage network includes the following steps:

[0008] A. Construct a deep residual shrinkage network, stack a certain number of adaptive slope module based residual shrinkage units (RSBU-ASB), input layer, convolution layer, batch normalization (BN), ReLU activation function, global average pooling (GAP) and cross entropy error function together to build a complete deep residual shrinkage network;

[0009] A1. Insert the soft threshold function as a nonlinear transformation layer into the deep residual shrinkage network structure to filter out the influence of noise-related features;

[0010] A2. Introduce the attention mechanism into the network and add an adaptive slope module to automatically obtain the optimal threshold for seismic signal denoising and achieve the best processing of different signals and different degrees of noise;

[0011] B. Establish an earthquake detection model to accurately obtain the detection effect of earthquake events.

[0012] Furthermore, in step A, a seismic signal of size C×W×1 is input through the input layer, where C is the number of channels, W is the width, and 1 is the height of the seismic signal (since the input is a three-component seismic waveform signal, the height is always 1); then a convolution layer is used to extract the features of each channel of the signal. In order to prevent overfitting and increase the nonlinear expression ability of the model, BN and ReLU activation functions are added before the convolution operation.

[0013] Furthermore, in step A, the convolution layer extracts the features of the input object and multiplies the output feature map of the previous layer by the convolution kernel according to certain rules.

[0014]

[0015] in is the jth feature of the lth layer in the output feature map, is the i-th feature of layer l-1 in the input feature map, is the convolution kernel size between the jth feature of layer l and the ith feature of layer l-1, is the corresponding bias, * represents the convolution operation, M j It is used to calculate the j-th feature.

[0016] Furthermore, in the step A, after extracting the features of the input object through the convolution layer, the extracted data is normalized through batch normalization to improve the convergence speed and generalization ability of the convolutional neural network, and a stretch parameter γ and an offset parameter β are introduced to restore the feature distribution learned by the network layer;

[0017] in

[0018]

[0019] In the formula, m is the batch size of samples, x i and i are the i-th input feature and output feature in the sample respectively.

[0020] Furthermore, after the convolution operation, the soft threshold module is used to estimate the threshold of the soft threshold function, and an adaptive slope module (ASB) is added to form a residual shrinkage unit based on the adaptive slope module (RSBU-ASB). The attention mechanism is used to automatically infer the most appropriate slope of the signal and further correct the threshold.

[0021] Furthermore, when the soft threshold module is used to estimate the threshold, the absolute value of the feature map is firstly GAPed to obtain a one-dimensional vector, and then the one-dimensional vector is passed to the two-layer FC to obtain the scale parameter α c Finally, the Sigmoid function is used to scale the scale parameter to between (0,1). The scale parameter calculation formula is:

[0022]

[0023] Among them, z c is the characteristic of the c-th layer neuron, τ c The threshold is obtained by multiplying the scale parameter by the absolute value of the one-dimensional feature vector, that is,

[0024]

[0025] Among them, i, j, c are the width, height and channel of the feature map x respectively.

[0026] Furthermore, in the adaptive slope module, the attention mechanism is used to automatically infer the most appropriate slope of the signal, and its output is

[0027]

[0028] Where β is the adaptive slope factor, q c is the characteristic of the c-th layer neuron;

[0029] Finally, the output of the soft threshold module is modified according to the adaptive slope factor β to obtain the optimal soft threshold, that is:

[0030]

[0031] According to the best soft threshold obtained, features with absolute values ​​less than the threshold are deleted, while features with absolute values ​​greater than the threshold are moved closer to 0 to achieve the purpose of shrinkage.

[0032] Furthermore, in step B, the earthquake detection model established includes a training set, a validation set and a test set, and the samples are randomly distributed to the established training set, validation set and test set in the ratio of 8:1:1, which are used for model training, adjustment of model hyperparameters and testing of the model earthquake event detection effect, respectively.

[0033] Furthermore, in the training set, multiple iterative trainings are set, and after repeated training and tuning experiments, the error of the training results is minimized, the model converges, and the optimal model is obtained.

[0034] Furthermore, after the model is trained with the training set, the established earthquake detection model is verified with samples in the test set.

[0035] Compared with the prior art, the advantages of the present invention are as follows: the present invention establishes an earthquake detection model by constructing a deep residual shrinkage network, in which a soft threshold function is inserted into the deep network structure as a nonlinear transformation layer to effectively filter the influence of noise-related features. At the same time, an attention mechanism is introduced into the network, and ASB is added to automatically obtain the optimal threshold for earthquake signal denoising, so as to achieve the best processing of different signals and different degrees of noise, thereby achieving the purpose of accurate detection of earthquake events under strong background noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 This is the process intention of the low signal-to-noise ratio earthquake detection model based on the deep residual shrinkage network provided by the present invention.

[0038] Figure 2 It is a seismic event and noise signal data diagram of the first data source of the low signal-to-noise ratio earthquake detection model based on the deep residual shrinkage network provided by the present invention.

[0039] Figure 3 It is a data diagram of earthquake events and noise signals in a microseismic monitoring data set of the second data source of the low signal-to-noise ratio earthquake detection model based on the deep residual shrinkage network provided by the present invention.

[0040] Figure 4 It is a curve diagram of accuracy and loss function of the model training process of the low signal-to-noise ratio earthquake detection model based on the deep residual shrinkage network provided by the present invention.

[0041] Figure 5 is the confusion matrix of the STEAD test set.

[0042] Figure 6 It is the ROC curve of the STEAD test set.

[0043] Figure 7 is the confusion matrix of the microseismic monitoring data test set.

[0044] Figure 8 It is the ROC curve of the microseismic monitoring data test set. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.

[0046] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0047] In the description of the present invention, it should be noted that if the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside" and the like appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the accompanying drawings, or is the orientation or position relationship in which the product of the invention is usually placed when used. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0048] In addition, the terms “first”, “second”, “third”, etc., if used, are merely used to distinguish between the descriptions and should not be understood as indicating or implying relative importance.

[0049] In addition, the terms "horizontal", "vertical", "overhanging", etc. do not mean that the components are required to be absolutely horizontal or overhanging, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0050] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", "connection", etc. should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0051] It should be noted that, in the absence of conflict, the features in the embodiments of the present invention may be combined with each other.

[0052] like Figure 1 As shown, a low signal-to-noise ratio earthquake detection model based on a deep residual shrinkage network includes the following steps:

[0053] A. Construct a deep residual shrinkage network, stack a certain number of adaptive slope module based residual shrinkage units (RSBU-ASB), input layer, convolution layer, batch normalization (BN), ReLU activation function, global average pooling (GAP) and cross entropy error function together to build a complete deep residual shrinkage network;

[0054] First, a seismic signal of size C×W×1 is input through the input layer, where C is the number of channels, W is the width, and 1 is the height of the seismic signal (since the input is a three-component seismic waveform signal, the height is always 1); then the convolution layer is used to extract the features of each channel of the signal. In order to prevent overfitting and increase the nonlinear expression ability of the model, BN and ReLU activation functions are added before the convolution operation.

[0055] The convolution layer extracts the features of the input object and multiplies the output feature map of the previous layer by the convolution kernel according to certain rules.

[0056]

[0057] in is the jth feature of the lth layer in the output feature map, is the i-th feature of layer l-1 in the input feature map, is the convolution kernel size between the jth feature of layer l and the ith feature of layer l-1, is the corresponding bias, * represents the convolution operation, M j It is used to calculate the j-th feature.

[0058] After extracting the features of the input object through the convolution layer, the extracted data is normalized through batch normalization to improve the convergence speed and generalization ability of the convolutional neural network, and the stretch parameter γ and offset parameter β are introduced to restore the feature distribution learned by the network layer;

[0059] in

[0060]

[0061] In the formula, m is the batch size of samples, x i and i are the i-th input feature and output feature in the sample respectively.

[0062] A1. Insert the soft threshold function as a nonlinear transformation layer into the deep residual shrinkage network structure to filter out the influence of noise-related features.

[0063] Specifically, after the convolution operation, the soft threshold module is used to estimate the threshold of the soft threshold function, and an adaptive slope module (ASB) is added to form a residual shrinkage unit based on the adaptive slope module (RSBU-ASB). The attention mechanism is used to automatically infer the most appropriate slope of the signal and further correct the threshold.

[0064] When using the soft threshold module to estimate the threshold, first use GAP on the absolute value of the feature map to obtain a one-dimensional vector, and then pass the one-dimensional vector to the two-layer FC to obtain the scale parameter αc Finally, the Sigmoid function is used to scale the scale parameter to between (0,1). The scale parameter calculation formula is:

[0065]

[0066] Among them, z c is the characteristic of the c-th layer neuron, τ c The threshold is obtained by multiplying the scale parameter by the absolute value of the one-dimensional feature vector, that is,

[0067]

[0068] Among them, i, j, c are the width, height and channel of the feature map x respectively.

[0069] A2. Introduce the attention mechanism into the network and add an adaptive slope module to automatically obtain the optimal threshold for seismic signal denoising and achieve the best processing of different signals and different degrees of noise;

[0070] In the adaptive slope module, the attention mechanism is used to automatically infer the most appropriate slope of the signal, and its output is

[0071]

[0072] Where β is the adaptive slope factor, q c is the characteristic of the c-th layer neuron;

[0073] Finally, the output of the soft threshold module is modified according to the adaptive slope factor β to obtain the optimal soft threshold, that is:

[0074]

[0075] According to the best soft threshold obtained, features with absolute values ​​less than the threshold are deleted, while features with absolute values ​​greater than the threshold are moved closer to 0 to achieve the purpose of shrinkage.

[0076] B. Establish an earthquake detection model to accurately obtain the detection effect of earthquake events.

[0077] In step B, the earthquake detection model established includes a training set, a validation set and a test set. The samples are randomly distributed to the established training set, validation set and test set in a ratio of 8:1:1, which are used for model training, adjustment of model hyperparameters and testing of the model earthquake event detection effect, respectively.

[0078] In the training set, multiple iterative trainings are set up. After repeated training and tuning experiments, the error of the training results is minimized, the model converges, and the optimal model is obtained.

[0079] After the model training is completed, the samples in the test set are used to verify the established earthquake detection model.

[0080] Specifically, if Figure 1 As shown, the present invention stacks a certain number of residual shrinkage units (RSBU-ASB), convolutional layers, batch normalization (BN), activation functions (ReLU), global average pooling (GAP) and fully connected layers (FC) together to obtain a complete deep residual shrinkage network. In this network, first, the input layer is used to input a three-component (i.e., channel C=3, width W=6000, height H=1) seismic signal with a length of 6,000; then a convolutional layer is used to set the number of channels to 12, the convolution kernel size to 3, and the step size to 1; then there are 5 combinations of residual shrinkage units (RSBU-ASB) (the number of channels is 12, 24, 48, 96 and 192, respectively), each combination contains two residual shrinkage units (RSBU-ASB), and the step sizes are 2 and 1, respectively; finally, GAP is used to reduce the feature map to a one-dimensional vector of 192×1, and a fully connected layer is used to realize the classification of seismic events and noise signals. The deep residual shrinkage network is based on the residual network. It inserts the soft threshold function as a nonlinear transformation layer into the deep network structure and applies a separate soft threshold to each channel of the feature map to remove noise-related features, so as to improve the model's ability to learn seismic event characteristics from low signal-to-noise ratio seismic signals, and ultimately achieve the purpose of accurate detection of seismic events under strong background noise.

[0081] When establishing the earthquake detection model, the first data source is the Stanford University Earthquake Dataset (STEAD), which consists of more than 1 million data (including earthquake events and noise signals) recorded by earthquake stations around the world from January 1984 to August 2018. Each earthquake waveform data in this data set is a 60s three-component signal with a sampling rate of 100Hz. The present invention randomly selects 75,000 earthquake events and noise signals to form an experimental data set, and randomly divides them into training set, validation set and test set at a ratio of 80%, 10%, and 10%, which are used for model training, model hyperparameter adjustment and model earthquake event detection effect testing respectively. The earthquake events and noise signals in the experimental data set are as follows: Figure 2 shown.

[0082] like Figure 3As shown, data source two uses microseismic monitoring data from a shale gas development area in southern Sichuan, China. This experiment aims to address the earthquake problem induced by hydraulic fracturing in the shale gas extraction process. 13 three-component detectors were deployed to collect earthquake records from February 2017 to July 2018. Each seismic waveform data in this data set is a 60s three-component signal with a sampling rate of 250Hz. Similarly, the present invention also randomly selects 75,000 seismic events and noise signals to form an experimental data set, and divides them into training set, validation set, and test set at a ratio of 80%, 10%, and 10%. From Figure 3 It can be seen that compared with natural seismic signals, the signal-to-noise ratio of microseismic signals generated by hydraulic fracturing is significantly lower, and microseismic events with weak energy are almost submerged in the strong background noise.

[0083] like Figure 4 As shown in , the model is trained according to the selected training data. After repeated training and tuning experiments, the error of the training result is minimized, the model converges, and the optimal model is obtained. The accuracy and loss function curves during the model training process are shown in Figure 4 As shown in the figure. As the number of iterations increases, the model converges at a faster speed, the training and validation loss function values ​​gradually decrease, and the accuracy gradually increases. When the iteration reaches 80 times, the model's loss function value no longer decreases, and the accuracy no longer increases, which indicates that the model tends to converge. The training and validation loss function values ​​of the final model stabilize at 0.390 and 0.392, and the corresponding accuracy stabilizes at 99.90% and 99.78%.

[0084] Then, the trained model was used to test 15,000 samples from data source 1 that were not involved in model training and verification, and the detection effect of earthquake events was obtained. The results were compared with the mainstream detection methods, and the results are shown in the following table:

[0085]

[0086] Table 1 Comparison of earthquake event detection results based on STEAD test data

[0087] And as Figure 5-Figure 6 As shown, it is clearly shown that the present invention uses the trained DRSN to correctly identify 14,862 samples, 136 earthquakes are misclassified as noise, and 2 noises are misclassified as earthquakes. Its accuracy, precision, recall and F1-score reach 99.08%, 99.97%, 98.19% and 99.07% respectively. It can be seen from various evaluation results that DRSN has a relatively optimal detection effect.

[0088] The trained model was used to test 15,000 samples from data source 2 that were not involved in model training and verification to detect microseismic events. The results were compared with mainstream detection methods, and are shown in the following table:

[0089]

[0090] Table 2 Comparison of detection results based on hydraulic fracturing microseismic monitoring data

[0091] It can be seen from Table 2 that the detection results of each deep learning method have the same trend as that of STEAD (experimental data 1). The earthquake event detection effects of each deep learning method are better than those of traditional STA / LTA, and DRSN has the best detection effect. 14,314 samples were correctly identified, 437 earthquake events were misclassified as noise signals, and 249 noise signals were misclassified as earthquake events. Its accuracy, precision, recall and F1-score reached 95.43%, 96.59%, 94.17% and 95.37% respectively, which has a relatively optimal detection effect.

[0092] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A low signal-to-noise ratio earthquake detection model based on a deep residual shrinkage network, characterized by: The following steps are involved: A. Construct a deep residual shrinkage network, stack a certain number of adaptive slope module based residual shrinkage units (RSBU-ASB), input layer, convolution layer, batch normalization (BN), ReLU activation function, global average pooling (GAP) and cross entropy error function together to build a complete deep residual shrinkage network; A1. Insert the soft threshold function as a nonlinear transformation layer into the deep residual shrinkage network structure to filter out the influence of noise-related features; A2. Introduce the attention mechanism into the network and add an adaptive slope module to automatically obtain the optimal threshold for seismic signal denoising and achieve the best processing of different signals and different levels of noise; B. Establish an earthquake detection model to accurately obtain the detection effect of earthquake events.

2. The low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network according to claim 1, characterized in that: In step A, a seismic signal of size C×W×1 is input through the input layer, where C is the number of channels, W is the width, and 1 is the height of the seismic signal (since the input is a three-component seismic waveform signal, the height is always 1); then a convolutional layer is used to extract the features of each channel of the signal. In order to prevent overfitting and increase the nonlinear expression ability of the model, BN and ReLU activation functions are added before the convolution operation.

3. The low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network according to claim 1, characterized in that: In step A, the convolution layer extracts the features of the input object and multiplies the output feature map of the previous layer by the convolution kernel according to certain rules. in is the jth feature of the lth layer in the output feature map, is the i-th feature of layer l-1 in the input feature map, is the convolution kernel size between the jth feature of layer l and the ith feature of layer l-1, is the corresponding bias, * represents the convolution operation, M j It is used to calculate the j-th feature.

4. The low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network according to claim 3 is characterized in that: In the step A, after extracting the features of the input object through the convolution layer, the extracted data is normalized through batch normalization to improve the convergence speed and generalization ability of the convolutional neural network, and the stretch parameter γ and the offset parameter β are introduced to restore the feature distribution learned by the network layer; wherein In the formula, m is the batch size of samples, x i and i are the i-th input feature and output feature in the sample respectively.

5. The low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network according to claim 4, characterized in that: After the convolution operation, the soft threshold module is used to estimate the threshold of the soft threshold function, and an adaptive slope module (ASB) is added to form a residual shrinkage unit based on the adaptive slope module (RSBU-ASB). The attention mechanism is used to automatically infer the most appropriate slope of the signal and further correct the threshold.

6. The low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network according to claim 5, characterized in that: When using the soft threshold module to estimate the threshold, first use GAP on the absolute value of the feature map to obtain a one-dimensional vector, and then pass the one-dimensional vector to the two-layer FC to obtain the scale parameter α c Finally, the Sigmoid function is used to scale the scale parameter to between (0,1). The scale parameter calculation formula is: Among them, z c is the characteristic of the c-th layer neuron, τ c The threshold is obtained by multiplying the scale parameter by the absolute value of the one-dimensional feature vector, that is, Among them, i, j, c are the width, height and channel of the feature map x respectively.

7. The low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network according to claim 1, characterized in that: In the adaptive slope module, the attention mechanism is used to automatically infer the most appropriate slope of the signal, and its output is Where β is the adaptive slope factor, q c is the characteristic of the c-th layer neuron; Finally, the output of the soft threshold module is modified according to the adaptive slope factor β to obtain the optimal soft threshold, that is: According to the best soft threshold obtained, features with absolute values ​​less than the threshold are deleted, while features with absolute values ​​greater than the threshold are moved closer to 0 to achieve the purpose of shrinkage.

8. The low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network according to claim 1, characterized in that: In step B, the earthquake detection model established includes a training set, a validation set and a test set. The samples are randomly distributed to the established training set, validation set and test set in a ratio of 8:1:1, which are used for model training, adjustment of model hyperparameters and testing of the model earthquake event detection effect, respectively.

9. The low signal-to-noise ratio earthquake detection model based on deep residual shrinkage network according to claim 8, characterized in that: In the training set, multiple iterative trainings are set up. After repeated training and tuning experiments, the error of the training results is minimized, the model converges, and the optimal model is obtained. After the model is trained with the training set, the established earthquake detection model is verified on the samples in the test set.

Images