A method for earthquake event detection based on CDIL-CNN combined with convolutional attention mechanism

Abstract

The present invention relates to the technical field of earthquake event detection, and is a method for earthquake event detection based on the combination of CDIL-CNN and convolutional attention mechanism, comprising: using a three-component earthquake data set to construct a network model, the network model includes a convolution combination module, an information fusion module, a convolutional attention module, a flattening layer, a first fully connected layer, a third excitation layer, a random hidden neuron layer, a second fully connected layer, and a fourth excitation layer; there are three convolution combination modules, and a circular dilated convolution module is used to perform dilated convolution operations to extract features of the three component earthquake data respectively. The information fusion module performs weighted average calculation on the output of the convolution combination module according to the minimum mean square error function to obtain preliminary detection results; the convolutional attention module performs weight adjustment. The present invention can efficiently and accurately complete earthquake event detection without manual labeling and modification of parameters and thresholds after network training.

Description

Technical Field

[0001] The present invention relates to the technical field of earthquake event detection, and in particular to an earthquake event detection method based on the combination of CDIL-CNN and convolutional attention mechanism. Background Art

[0002] As the earthquake detection network continues to improve, seismic stations have recorded a large amount of waveform data, which provides data support for earthquake location, analysis and other research. Seismic data is an important source of information for geological condition analysis and has important scientific value for studying earthquake occurrence mechanisms and earthquake early warning. Through the analysis of seismic data, we can better understand the laws of earthquake activity, predict potential earthquake disasters, and effectively reduce the impact of geological disasters. At present, there are three main ways to detect effective earthquake events from a large amount of waveform data: manual annotation, traditional methods, and deep learning methods.

[0003] Manually labeling a large amount of seismic data requires a lot of time and effort. Traditional seismic event detection methods mainly include the long-short time window ratio method (STA / LTA), the energy ratio method (Energy Ratio, ER) and the long-short time window indicator ratio method (SLTA). These traditional seismic event detection methods all set a threshold based on changes in the energy or amplitude of the seismic waveform. Improper threshold setting may result in missed or false detections, and these methods are difficult to accurately distinguish between noise and seismic events in low signal-to-noise ratio seismic waveform data detection. Summary of the invention

[0004] In view of the fact that it is difficult to accurately distinguish noise and seismic events in the detection of low signal-to-noise ratio seismic waveform data in the prior art, the present invention proposes a seismic event detection method based on the combination of CDIL-CNN and convolutional attention mechanism.

[0005] The present invention is achieved in this way.

[0006] A method for earthquake event detection based on CDIL-CNN combined with convolutional attention mechanism, including:

[0007] S1: Use three-component seismic data samples, perform normalization processing, and annotate the long-sequence seismic waveform data of each three-component seismic data sample with noise and seismic events according to the arrival time and the end time of the wave, and then divide it into a training set and a test set;

[0008] S2: Construct a network model, which includes a convolution combination module, an information fusion module, a convolution attention module, a flattening layer, a first fully connected layer, a third excitation layer, a random hidden neuron layer, a second fully connected layer, and a fourth excitation layer in sequence from input to output; the convolution combination module is divided into three groups, each group includes a circular dilated convolution module, a first excitation layer, a batch normalization layer, a second excitation layer, and a maximum pooling layer: wherein the probability of selecting the random hidden neuron layer is 0.4, the circular dilated convolution module in each group includes 4 layers of circular dilated convolution layers for dilated convolution operations, and the circular dilated convolution modules in the three groups respectively complete the feature extraction of the three-component seismic data; the information fusion module uses the mean square error function as the weight adjustment function to obtain a preliminary detection result; the convolution attention module weights the signal-to-noise ratio of different component signals according to the signal-to-noise ratio of the three component signals of the three-component seismic data with the same arrival time and end time, the preliminary detection result of the output of the information fusion module, and the true value of the label annotated by the seismic event;

[0009] S3: Use the training set in S1 to train the network model. After the training is completed, save the network model. After testing with the test set, input the three-component earthquake data to be processed into the trained network model for earthquake event detection, and output the detection result of the earthquake event.

[0010] Furthermore, the numbers of hidden channels corresponding to the circular dilated convolution modules in the three groups from the first circular dilated convolution module to the third circular dilated convolution module are 64, 128, and 256, respectively, wherein the parameters of each circular dilated convolution module are set to an initial input channel number of 1, and the normalized three-component seismic data are input into the circular dilated convolution modules with hidden channels of 64, 128, and 256 according to the signal-to-noise ratio from small to large to perform dilated convolution operations to varying degrees; each circular dilated convolution module includes 4 circular dilated convolution layers; wherein the initial input channel number of the first circular dilated convolution layer of each circular dilated convolution module is the initial input channel number of the corresponding circular dilated convolution module, and starting from the second circular dilated convolution layer, the input channel number is the hidden channel number corresponding to the circular dilated convolution module; setting the first circular dilated convolution layer to the hidden channel number corresponding to the circular dilated convolution module, the first circular dilated convolution layer is set to the hidden channel number corresponding to the circular dilated convolution module; setting the first circular dilated convolution layer to the hidden channel number corresponding to the circular dilated convolution module, the first circular dilated convolution layer is set to the hidden channel number corresponding to the circular dilated convolution module; The dilation rate of the circular dilated convolutional layer is , The values ​​are 1, 2, 3, and 4 respectively.

[0011] Furthermore, each circular dilated convolution module uses the information of the previous and next elements in the sequence formed by the three-component seismic data to fill in the convolution operation. The filling mode is cyclic filling, and the filling size is determined by the dilation rate and the convolution kernel size. The formula is:

[0012] ,

[0013] in, Represents the fill size, Represents the current expansion rate, represents the convolution kernel size, To round the calculation result, remove the decimal points.

[0014] Furthermore, the operation process from the first circular dilated convolution layer to the fourth circular dilated convolution layer in each circular dilated convolution module is:

[0015] The first circular dilated convolutional layer:

[0016] ,

[0017] The second circular dilated convolutional layer to the fourth circular dilated convolutional layer:

[0018] ,

[0019] in, Representative Quantity Channel No. The output of the circular dilated convolution layer, Representative Quantity Channel No. The output of the circular dilated convolution layer, is the earthquake sequence index value, ranging from 0 to 5999. Take 1 to , Indicates The number of hidden channels of the circular dilated convolution module, Take 1, 2, 3, Representative Layer convolution kernel size, Take 1, 2, 3, 4, For the The weight coefficient corresponding to the channel convolution kernel, Is the convolution kernel coefficient index, ranging from 0 to , Representative Layer fill size, Representative The size of the layer expansion rate, Represents the first component of the input three-component seismic data Component waveform data, The values ​​are 1, 2, and 3, and the length of the three-component seismic data sequence is =6000, Represents circular filling. If If it is a negative number, the data at the end of the sequence will be added;

[0020] After The data after the output of the circular dilated convolution module for:

[0021] ,

[0022] The feature dimension is ( , 6000).

[0023] Furthermore, the information fusion module uses a mean square error function as a weight adjustment function to obtain a preliminary detection result, including:

[0024] The mean square error function used is: ,

[0025] The characteristics of the three component seismic data are weighted averaged by the mean square error function:

[0026] ,

[0027] by and Respectively and To update, The value is [0.01,0.001], is the length of the three-component seismic data sequence, For the The mean square error between the initial prediction results of the components and the labels, initial Set to uniform distribution 1* Weight matrix, initial is a 1*6000 bias matrix with all zeros; is the true value of the label, The three-component seismic data The fusion output of the components is The value range is 0-5999. for The elements; For the After the output of the circular dilated convolution module, the dimension is *6000 data.

[0028] Furthermore, the convolutional attention module weight calculation formula is as follows:

[0029] ,

[0030] in, for The sum of the differences from the true value of the label, The value range is from 0 to 5999;

[0031] ,

[0032] For the The initial weights of the components are calculated based on the initial prediction errors;

[0033] ,

[0034] in For the The initial weight distribution of each component is calculated based on the prediction error. represents the signal-to-noise ratio of the three components of seismic data, Based on the initial weight distribution and The final three-component weight distribution obtained by size calculation;

[0035] ,

[0036] The output value after assigning weights to the convolutional attention module.

[0037] Furthermore, the loss function of the network model adopts the cross entropy loss function for:

[0038] ,

[0039] in, and are the true value of the label and the predicted value of the label, respectively. is the earthquake sequence index value, is the length of the three-component seismic data sequence.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention utilizes three circular dilated convolution modules to respectively realize feature extraction and learning of three components of three-component seismic data. Different degrees of dilated convolution are adopted for three components with different signal-to-noise ratios. Circular dilated convolution uses symmetric convolution to mix information, ensuring that each position can obtain feature information of different positions in the sequence in a cyclic manner. In addition, the dilated convolution method and the dilation rate variation formula are used to avoid excessive learning of local noise and seismic event features due to too fast expansion of the dilation rate during the dilated convolution process. Seismic data features can be extracted and learned more efficiently, and seismic event detection can be completed accurately. In order to enable the convolution combination module and the convolution attention module to play a better role, an information fusion module is designed between them. The results of the dilated convolution of the convolution combination module are weighted averaged based on the mean square error function to obtain a preliminary detection result. According to the difference in signal-to-noise ratio of the three component signals of the three-component seismic data with the same wave arrival time and end time, the designed network model combines the convolutional attention module to adaptively adjust the weight of each component and fuse them for subsequent network training. In this module, we input the data after the information fusion module, the signal-to-noise ratio information of the three-component seismic data, and the label information respectively, and then adjust the weight according to the acquired feature information and the signal-to-noise ratio of different components to achieve the purpose of accurately detecting earthquake events;

[0042] Compared with manual labeling and traditional methods, the present invention can efficiently and accurately complete earthquake event detection without the need for manual labeling and modification of parameters and thresholds after network training. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of a method provided by an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the network structure of a network model provided by an embodiment of the present invention;

[0045] Figure 3 A schematic diagram of the network structure of a circular dilated convolution module provided in an embodiment of the present invention;

[0046] Figure 4 A diagram of a four-layer dilated convolution process of a circular dilated convolution module provided in an embodiment of the present invention; DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0048] See also Figure 1 As shown, a method for earthquake event detection based on CDIL-CNN combined with convolutional attention mechanism includes:

[0049] S1: A three-component seismic data sample is used, normalized, and noise and seismic events are marked. Then, each seismic event band of the same length is subdivided into seismic events and noise according to the arrival time and the end time, and divided into a training set and a test set; the event marking here is to mark all effective seismic event bands as 1 and the noise band as 0 according to the arrival time and the end time of the waveform in each three-component seismic data sample, and divide them into a training set and a test set according to a ratio of 4:1; the three-component seismic data sample in this embodiment adopts the Stanford public three-component seismic data set, and selects 30,000 groups of 0-30db three-component seismic data with a length of 60s and a sampling rate of 100Hz;

[0050] S2: Build a network model, see Figure 2 Combination Figure 3 , Figure 4 As shown, the network model includes a convolution combination module, an information fusion module, a convolution attention module, a flattening layer, a first fully connected layer, a third excitation layer, a random hidden neuron layer, a second fully connected layer, and a fourth excitation layer in sequence from input to output; the convolution combination module is divided into three groups, each group includes a circular dilated convolution module, a first excitation layer, a batch normalization layer, a second excitation layer, and a maximum pooling layer: wherein, the probability of selecting the random hidden neuron layer is 0.4, and according to the feature requirements of the three-component seismic data, the circular dilated convolution module in each group includes 4 layers of circular dilated convolution layers for dilated convolution operations, and the circular dilated convolution modules in the three groups respectively complete the feature extraction of the three-component seismic data. ; Between the designed convolution combination module and the convolution attention module, an information fusion module is designed for transition connection. The information fusion module uses the mean square error function as the weight adjustment function to capture the key information of the wave arrival time and the wave arrival end time to obtain a preliminary detection result; according to the difference in signal-to-noise ratio of the three component signals of the three-component seismic data with the same wave arrival time and end time, the convolution attention module adjusts the weights of the signal-to-noise ratio of different component signals according to the signal-to-noise ratio of the three component signals of the three-component seismic data with the same wave arrival time and end time, the preliminary detection results output by the information fusion module, and the true value of the label annotated by the earthquake event;

[0051] The three circular dilated convolution modules in the three groups have corresponding hidden channel numbers of 64, 128, and 256 from the first to the third, respectively. The parameters of each circular dilated convolution module are set to an initial input channel number of 1. The setting of the convolution kernel size is based on the proprietary characteristics of seismic data. A larger convolution kernel is used in the shallow layer of the circular dilated convolution module to capture the propagation characteristics of long-distance seismic waves. The initial convolution kernel size is set to 7. As the number of layers increases, in order to obtain local information of seismic signals to separate noise and effective events, but at the same time, it is also necessary to avoid the loss of key information due to too small a convolution kernel, so the size is set. , Represents the minimum convolution kernel, and the size of the convolution kernel varies according to the number of layers. . represents the convolution kernel size, represents the convolution kernel size of the previous layer, and the number of layers is set to 4, that is, each circular dilated convolution module contains 4 circular dilated convolution layers, such as Figure 3 As shown in the figure, the convolution kernel size of the first circular dilated convolution layer is 7, the convolution kernel size of the second circular dilated convolution layer is 5, the convolution kernel size of the third circular dilated convolution layer is 3, and the convolution kernel size of the fourth circular dilated convolution layer is 3; the initial input channel number of the first circular dilated convolution layer of each circular dilated convolution module is the initial input channel number of the circular dilated convolution module, and starting from the second circular dilated convolution layer, the number of input channels is the corresponding number of hidden channels. In the earthquake event detection task, dilated convolution is used to expand the receptive field and improve the detection accuracy of the separation moment of noise and earthquake events. However, it is also necessary to avoid the expansion rate expanding too fast during the dilation convolution process and over-learning the local noise and earthquake event characteristics. Therefore, the first The dilation rate of the circular dilated convolutional layer is , The values ​​are 1, 2, 3, and 4 in sequence; that is, the expansion rate of the first circular dilated convolutional layer 1, the expansion rate of the second circular expansion convolution layer 1, the expansion rate of the third circular expansion convolution layer 3, the fourth circular dilated convolutional layer The expansion rate is 5; the final output is the result of multiple convolutions.

[0052] In order to avoid the loss of boundary information of the three-component seismic data during the convolution operation, each circular dilated convolution module uses the information of the previous and next elements in the sequence formed by the three-component seismic data to fill the data. The filling mode is cyclic filling, and the filling size is determined by the dilation rate and the convolution kernel size. The formula is:

[0053] ,

[0054] in, Represents the fill size, Represents the current expansion rate, represents the convolution kernel size, To round the calculation result, remove the decimal points.

[0055] The operation process from the first circular dilated convolution layer to the fourth circular dilated convolution layer in each circular dilated convolution module is as follows:

[0056] The first circular dilated convolutional layer:

[0057] ,

[0058] The second circular dilated convolutional layer to the fourth circular dilated convolutional layer:

[0059] ,

[0060] in, Representative Quantity Channel No. The output of the circular dilated convolution layer, is the earthquake sequence index value, ranging from 0 to 5999. Take 1 to , Indicates The number of hidden channels of the circular dilated convolution module, Take 1, 2, 3, Representative Layer convolution kernel size, Take 1, 2, 3, 4, For the The weight coefficient corresponding to the channel convolution kernel, Is the convolution kernel coefficient index, ranging from 0 to , Representative Layer fill size, Representative The size of the layer expansion rate, Represents the first component of the input three-component seismic data Component waveform data, The values ​​are 1, 2, and 3, and the length of the three-component seismic data sequence is =6000, Represents circular filling. If If it is a negative number, the data at the end of the sequence will be added;

[0061] After The data after the output of the circular dilated convolution module for:

[0062] ,

[0063] The feature dimension is ( , 6000).

[0064] The information fusion module uses the mean square error function as the weight adjustment function to obtain a preliminary detection result, including:

[0065] The mean square error function used is: ,

[0066] The characteristics of the three component seismic data are weighted averaged by the mean square error function:

[0067] ,

[0068] by and Respectively and To update, The value is [0.01,0.001], is the length of the three-component seismic data sequence, For the The mean square error between the initial prediction results of the components and the labels, initial Set to uniform distribution 1* Weight matrix, initial is a 1*6000 bias matrix with all zeros; is the true value of the label, The three-component seismic data The fusion output of the components is The value range is from 0 to 5999; for The elements, For the After the output of the circular dilated convolution module, the dimension is *6000 data.

[0069] The weight calculation formula of the convolutional attention module is as follows:

[0070] ,

[0071] in, for The sum of the differences from the true value of the label, is the earthquake sequence index value, is the length of the three-component seismic data sequence.

[0072] ,

[0073] in for The sum of the differences between the actual values ​​of the labels represents the current model The initial prediction error of the component, For the The initial weights of the components are calculated based on the initial prediction errors.

[0074] ,

[0075] in For the The initial weight distribution of each component is calculated based on the prediction error. represents the signal-to-noise ratio of the three components of seismic data, Based on the initial weight distribution and The size calculation obtains the final three-component weight distribution.

[0076] ,

[0077] in The output value after the convolutional attention mechanism assigns weights.

[0078] The loss function of the network model adopts the cross entropy loss function for:

[0079] ,

[0080] in, and are the true value of the label and the predicted value of the label, respectively. is the earthquake sequence index value, is the length of the three-component seismic data sequence.

[0081] At the end of the network model, the Sigmoid function in the fourth excitation layer is used to output the noise, effective events, and noise probability distribution corresponding to each sample band in earthquake event detection.

[0082] Finally, the training set in S1 is used to train the network model. After the training is completed, the network model is saved. After testing with the test set, the three-component earthquake data to be processed is input into the trained network model for earthquake event detection, and the output is the detection result of the earthquake event.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for earthquake event detection based on CDIL-CNN combined with convolutional attention mechanism, characterized in that: include: S1: Use three-component seismic data samples, perform normalization processing, and annotate the long-sequence seismic waveform data of each three-component seismic data sample with noise and seismic events according to the arrival time and the end time of the wave, and then divide it into a training set and a test set; S2: Construct a network model, which includes a convolution combination module, an information fusion module, a convolution attention module, a flattening layer, a first fully connected layer, a third excitation layer, a random hidden neuron layer, a second fully connected layer, and a fourth excitation layer in sequence from input to output; the convolution combination module is divided into three groups, each group includes a circular dilated convolution module, a first excitation layer, a batch normalization layer, a second excitation layer, and a maximum pooling layer: wherein the probability of selecting the random hidden neuron layer is 0.4, the circular dilated convolution module in each group includes 4 layers of circular dilated convolution layers for dilated convolution operations, and the circular dilated convolution modules in the three groups respectively complete the feature extraction of the three-component seismic data; the information fusion module uses the mean square error function as the weight adjustment function to obtain a preliminary detection result; the convolution attention module weights the signal-to-noise ratio of different component signals according to the signal-to-noise ratio of the three component signals of the three-component seismic data with the same arrival time and end time, the preliminary detection result of the output of the information fusion module, and the true value of the label annotated by the seismic event; S3: Use the training set in S1 to train the network model. After the training is completed, save the network model. After testing with the test set, input the three-component earthquake data to be processed into the trained network model for earthquake event detection, and output the detection result of the earthquake event.

2. The earthquake event detection method based on CDIL-CNN combined with convolutional attention mechanism according to claim 1, characterized in that: The numbers of hidden channels corresponding to the circular dilated convolution modules in the three groups from the first circular dilated convolution module to the third circular dilated convolution module are 64, 128, and 256 respectively, where the parameters of each circular dilated convolution module are set to an initial input channel number of 1, and the normalized three-component seismic data are input into the circular dilated convolution modules with hidden channels of 64, 128, and 256 according to the signal-to-noise ratio from small to large to perform dilated convolution operations to varying degrees; each circular dilated convolution module includes 4 circular dilated convolution layers; the initial input channel number of the first circular dilated convolution layer of each circular dilated convolution module is the initial input channel number of the corresponding circular dilated convolution module, and starting from the second circular dilated convolution layer, the input channel number is the hidden channel number corresponding to the circular dilated convolution module; setting the first circular dilated convolution layer to the first circular dilated convolution layer, the input channel number is the hidden channel number corresponding to the circular dilated convolution module; setting the first circular dilated convolution layer to the first circular dilated convolution layer, the input channel number is the hidden channel number corresponding to the circular dilated convolution module; setting the first circular dilated convolution layer to the first circular dilated convolution layer, the input channel number is the hidden channel number corresponding to the circular dilated convolution module. The dilation rate of the circular dilated convolutional layer is , The values ​​are 1, 2, 3, and 4 respectively.

3. The earthquake event detection method based on CDIL-CNN combined with convolutional attention mechanism according to claim 1, characterized in that: During the convolution operation, each circular dilated convolution module uses the information of the previous and next elements in the sequence formed by the three-component seismic data to fill the data. The filling mode is cyclic filling, and the filling size is determined by the dilation rate and the convolution kernel size. The formula is: ， in, Represents the fill size, Represents the current expansion rate, represents the convolution kernel size, To round the calculation result, remove the decimal points.

4. The earthquake event detection method based on CDIL-CNN combined with convolutional attention mechanism according to claim 3 is characterized in that: The operation process from the first circular dilated convolution layer to the fourth circular dilated convolution layer in each circular dilated convolution module is: The first circular dilated convolutional layer: ， From the second circular dilated convolutional layer to the fourth circular dilated convolutional layer: ， in, Representative Quantity Channel No. The output of the circular dilated convolution layer, Representative Quantity Channel No. The output of the circular dilated convolution layer, is the earthquake sequence index value, ranging from 0 to 5999. Take 1 to , Indicates The number of hidden channels of the circular dilated convolution module, Take 1, 2, 3, Representative Layer convolution kernel size, Take 1, 2, 3, 4, For the The weight coefficient corresponding to the channel convolution kernel, Is the convolution kernel coefficient index, ranging from 0 to , Representative Layer fill size, Representative The size of the layer expansion rate, Represents the first component of the input three-component seismic data Component waveform data, The values ​​are 1, 2, and 3, and the length of the three-component seismic data sequence is =6000, Represents circular filling. If If it is a negative number, the data at the end of the sequence will be added; After The data after the output of the circular dilated convolution module for: ， The feature dimension is ( , 6000).

5. The earthquake event detection method based on CDIL-CNN combined with convolutional attention mechanism according to claim 4 is characterized in that: The information fusion module uses a mean square error function as a weight adjustment function to obtain a preliminary detection result, including: The mean square error function used is: , The characteristics of the three component seismic data are weighted averaged by the mean square error function: ， by and Respectively and To update, The value is [0.01,0.001], is the length of the three-component seismic data sequence, For the The mean square error between the initial prediction results of the components and the labels, initial Set to uniform distribution 1* Weight matrix, initial is a 1*6000 bias matrix with all zeros; is the true value of the label, The three-component seismic data The fusion output of the components is The value range is 0-5999. for The elements.

6. The earthquake event detection method based on CDIL-CNN combined with convolutional attention mechanism according to claim 5, characterized in that: The weight calculation formula of the convolutional attention module is as follows: ， in, for The sum of the differences from the true value of the label; ， For the The initial weights of the components are calculated based on the initial prediction errors; ， in For the The initial weight distribution of each component is calculated based on the prediction error. represents the signal-to-noise ratio of the three components of seismic data, Based on the initial weight distribution and The final three-component weight distribution obtained by size calculation; ， The output value after assigning weights to the convolutional attention module.

7. The earthquake event detection method based on CDIL-CNN combined with convolutional attention mechanism according to claim 1, characterized in that: The loss function of the network model adopts the cross entropy loss function for: ， in, and are the true value of the label and the predicted value of the label, respectively. is the earthquake sequence index value, is the length of the three-component seismic data sequence.

Images