KR102323118B1 - Earthquake event classification method using attention-based convolutional neural network, recording medium and device for performing the method - Google Patents

Abstract

A seismic event classification method using a state-based neural network includes: centering and preprocessing input seismic data; extracting a feature map by non-linearly transforming the preprocessed seismic data through a plurality of convolutional layers of three or more; Measuring the importance of the learned feature based on the attention technique that models the interdependence between channels of the feature map on the nonlinearly transformed seismic data; correcting the feature value through element-by-element multiplication of the measured importance value with the learned feature map; down-sampling through max-pooling based on the feature value; and classifying earthquake events by normalizing the down-sampled feature values. Accordingly, through attention-based deep learning, key features inherent in large amounts of/complex data are extracted, thereby overcoming the limitations of existing micro-seismic sensing technologies and enabling seismic detection even in low SNR environments.

Description

A method for classifying seismic events using an attention-based neural network, and a recording medium and apparatus for performing the same

The present invention relates to a seismic event classification method using a caution-based neural network, a recording medium and an apparatus for performing the same, and more particularly, extracts major characteristic information inherent in data through attention-based deep learning, and through this It is about a technology that accurately classifies events such as various micro-earthquakes and artificial earthquakes.

Recently, natural disasters such as extreme weather are increasing due to climate change such as global warming.

In order to prepare for the increasing risk of earthquakes, various studies are being conducted to detect or predict the occurrence of earthquakes in advance and prepare for them. In particular, in the seismic event classification problem, there are problems of classifying not only strong earthquakes but also various (micro-earthquakes, artificial earthquakes) seismic events.

The STA/LTA method is a representative method of detecting earthquakes through the threshold ratio of Short Time Average (STA) and Long Time Average (LTA), and is not suitable in a low SNR environment. This method determines whether to detect earthquakes based on STA/LTA results from multiple stations.

The autocorrelation-based method detects earthquakes based on the correlation between repeated earthquake waveforms in a single area, and although it has excellent performance, it has limitations in its utility for overloading the computational amount and for long-time signals.

The template matching method is a technique proposed to improve the computational overload problem of the autocorrelation method. After setting an earthquake template, earthquakes are detected through correlation between the template and the input earthquake waveform.

The number of templates has a major impact on performance, and approaches to reduce the number of templates through PCA (Principle Component Analysis) or subspace learning have been proposed.

Recently, a Fingerprint And Similarity Thresholding (FAST) method has been proposed to improve the computational amount of template matching. This method is an unsupervised seismic detection method that extracts seismic features similar to fingerprints through spectral image and haar wavelet transformation, and reduces the amount of similarity measurement calculations through local sensitive hashing.

However, in the conventional techniques, the classification performance with noise is not satisfactory due to the low SNR environment for sensing small-scale earthquakes compared to large-scale earthquakes, and artificial earthquakes such as artificial blasting and nuclear tests detect natural small earthquakes. become a big hindrance.

KR 10-1914657 B1 KR 10- 1919098 B1

Accordingly, the technical problem of the present invention was conceived in this regard, and an object of the present invention is to provide a seismic event classification method using an attention-based neural network for building a robust earthquake detection system in a situation in which noise and artificial earthquakes are mixed. will be.

Another object of the present invention is to provide a recording medium in which a computer program for performing the seismic event classification method using the attention-based neural network is recorded.

Another object of the present invention is to provide an apparatus for performing a seismic event classification method using the attention-based neural network.

An earthquake event classification method using a caution-based neural network according to an embodiment of the present invention for realizing the object of the present invention, centering the input earthquake data and pre-processing; extracting a feature map by non-linearly transforming the preprocessed seismic data through a plurality of convolutional layers of three or more; Measuring the importance of the learned feature based on the attention technique that models the interdependence between channels of the feature map on the nonlinearly transformed seismic data; correcting the feature value through element-by-element multiplication of the measured importance value with the learned feature map; down-sampling through max-pooling based on the feature value; and classifying earthquake events by normalizing the down-sampled feature values.

In an embodiment of the present invention, measuring the importance of the learned feature may include performing a squeeze operation for compressing the feature map while maintaining global information in each channel.

In an embodiment of the present invention, the step of performing the squeeze operation includes using a global weight average pooling method that spreads a histogram of a low-contrast image distribution based on contrast stretching. can

In an embodiment of the present invention, the step of measuring the importance of the learned feature is adapted according to the importance of each channel using two fully-connected (FC) layers and a sigmoid activation function. It may further include the step of calculating an excitation (Excitation) for re-correcting the importance value.

In an embodiment of the present invention, the step of calculating the excitation may include: passing the compressed feature map through a dimensionality reduction layer having a preset reduction rate; and restoring the data passing through the dimension reduction layer to the channel dimension of the feature map.

In an embodiment of the present invention, the step of non-linearly mapping the input data of the preprocessed seismic data through three or more plurality of convolutional layers comprises: converting the preprocessed seismic data to the first convolutional layer and the last of the plurality of convolutional layers. The method may further include batch normalizing feature values in the convolutional layer.

In an embodiment of the present invention, in the batch normalizing of the feature values, the average and variance are obtained for each feature value and then normalized, and a scale factor and a shift factor for the normalized feature values. ) can be added to convert

In an embodiment of the present invention, classifying the seismic event by normalizing the down-sampled feature value includes passing first and second fully-connected (FC) layers, and the first fully-connected layer is A dropout normalization process that activates each neuron according to a probability value and applies it to learning may be performed.

In a computer-readable storage medium according to an embodiment for realizing another object of the present invention, a computer program for performing a seismic event classification method using an attention-based neural network is recorded.

An earthquake event classification apparatus using a caution-based neural network according to an embodiment for realizing another object of the present invention includes: a preprocessing unit for preprocessing input earthquake data by centering it; The preprocessed seismic data is non-linearly transformed to extract a feature map, and the non-linearly transformed seismic data is used to measure the importance of the learned feature based on the attention technique that models the interdependence between channels of the feature map, and the measured importance value is A feature comprising three or more convolutional layers that correct a feature value through element-by-element multiplication with the learned feature map, and down-sample through max-pooling based on the feature value extraction unit; and a classification unit including first and second fully-connected (FC) layers for classifying an earthquake event by normalizing the down-sampled feature value.

In an embodiment of the present invention, each convolutional layer of the feature extraction unit includes: a nonlinear transformation unit configured to nonlinearly transform the preprocessed earthquake data; an attention module that measures the importance of features learned based on an attention technique that models the interdependence between channels of the feature map using nonlinearly transformed seismic data and outputs a feature value; and a max pooling unit for down-sampling through max-pooling based on the feature value.

In an embodiment of the present invention, the first convolutional layer and the last convolutional layer of the feature extraction unit are subjected to normalization after obtaining the average and variance for each feature value, and a scale factor and a shift factor for the normalized feature values It may further include a batch normalizer that converts by adding a shift factor.

In an embodiment of the present invention, the attention module includes: a squeeze operation unit for compressing a feature map while maintaining global information in each channel of nonlinearly transformed seismic data; an excitation operation unit that adaptively re-corrects the importance value according to the importance of each channel using two fully-connected (FC) layers and a sigmoid activation function; and a scale unit that corrects the feature value through element-by-element multiplication of the measured importance value with the learned feature map.

In an embodiment of the present invention, the excitation calculating unit may include: a dimension reduction unit passing the compressed feature map through a dimensionality reduction layer having a preset reduction rate; and a dimension restoration unit that restores the data passing through the dimension reduction layer to the channel dimension of the feature map.

In an embodiment of the present invention, the first fully connected layer of the classifier may perform a dropout normalization process that activates each neuron according to a probability value and applies it to learning.

According to this seismic event classification method using an attention-based neural network, the main key features inherent in a large amount of/complex data are extracted through attention-based deep learning, thereby overcoming the limitations of the existing micro-earthquake detection technology. It enables seismic detection even in low SNR environments.

In addition, by providing accurate and prompt earthquake information, it can be expected to reduce anxiety about earthquakes and improve the reliability of warnings. Based on the seismic detection and analysis information, it is possible to induce construction in a safe place when planning the construction of major industrial facilities or to reduce the possibility of a large-scale disaster by preparing for earthquake-resistant design.

Furthermore, it is possible to derive areas of interest by analyzing accumulated micro-earthquake detection information, and it is possible to quickly alert and respond to earthquake damage by intensively monitoring nearby areas.

1 is a block diagram of an earthquake event classification apparatus using an attention-based neural network according to an embodiment of the present invention.
FIG. 2 is a diagram showing the CNN structure of the earthquake event classification apparatus using the attention-based neural network of FIG. 1 .
FIG. 3 is a block diagram illustrating first and eighth convolutional layers in the seismic event classification apparatus using the attention-based neural network of FIG. 1 .
FIG. 4 is a block diagram illustrating second to seventh convolutional layers in the seismic event classification apparatus using the attention-based neural network of FIG. 1 .
5 is a block diagram illustrating an attention module of the earthquake event classification apparatus using the attention-based neural network of FIG. 1 .
6 is a flowchart of a seismic event classification method using an attention-based neural network according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [0023] Reference is made to the accompanying drawings, which show by way of illustration specific embodiments in which the present invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention are different but need not be mutually exclusive. For example, certain shapes, structures, and characteristics described herein with respect to one embodiment may be embodied in other embodiments without departing from the spirit and scope of the invention. In addition, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the present invention. Accordingly, the detailed description set forth below is not intended to be taken in a limiting sense, and the scope of the invention, if properly described, is limited only by the appended claims, along with all scope equivalents to those claimed. Like reference numerals in the drawings refer to the same or similar functions throughout the various aspects.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

1 is a block diagram of an earthquake event classification apparatus using an attention-based neural network according to an embodiment of the present invention. FIG. 2 is a diagram showing the CNN structure of the earthquake event classification apparatus using the attention-based neural network of FIG. 1 .

The earthquake event classification device (10, hereinafter device) using an attention-based neural network according to the present invention extracts main characteristic information inherent in the data through attention-based deep learning, and through this, the limitations of the existing earthquake detection technology are overcome. want to overcome

Referring to FIG. 1 , an apparatus 10 according to the present invention includes a preprocessing unit 110 , a feature extraction unit 130 , and a classification unit 150 .

In the device 10 of the present invention, software (application) for performing seismic event classification using an attention-based neural network may be installed and executed, and the preprocessor 110, the feature extraction unit 130, and the classification The configuration of the unit 150 may be controlled by software for performing seismic event classification using the attention-based neural network running on the device 10 .

The device 10 may be a separate terminal or a module of the terminal. In addition, the configuration of the pre-processing unit 110 , the feature extracting unit 130 , and the classifying unit 150 may be formed as an integrated module or may consist of one or more modules. However, on the contrary, each configuration may be formed of a separate module.

The device 10 may be movable or stationary. The apparatus 10 may be in the form of a server or an engine, and may be a device, an application, a terminal, a user equipment (UE), a mobile station (MS), or a wireless device. (wireless device), may be called other terms such as a handheld device (handheld device).

The device 10 may execute or manufacture various software based on an operating system (OS), that is, the system. The operating system is a system program for software to use the hardware of the device, and is a mobile computer operating system such as Android OS, iOS, Windows Mobile OS, Bada OS, Symbian OS, Blackberry OS, and Windows series, Linux series, Unix series, It can include all computer operating systems such as MAC, AIX, and HP-UX.

First, a machine learning algorithm related to the present invention will be described. In the recognition method using a Neural Network (NN), a function y=f(x) that derives a result y suitable for the purpose from the training data x can be created using the neural network.

At this time, a plurality of variables exist in the function f(x), and the process of finding the plurality of variables through the learning data is called a learning process. The basic structure of a neural network is composed of an input layer - a hidden layer - an output layer, and the connection between layers is a fully connected structure. The process of obtaining a classification result y from new data x by applying the learned multiple variable values is called a test process.

In a convolutional neural network (CNN), when the dimension of learning data (eg, image input) increases during the learning process, the amount of parameters to be learned increases exponentially. A sufficient amount of data is required to learn these parameters, but in reality, the conditions are not met.

On the other hand, CNN has the advantage of sufficiently reducing the number of parameters to be learned through a convolution operation with characteristics of local connection and weight sharing. In addition, it has the advantage of automatically generating convolutional filters that are effective for the purpose through learning. CNN consists of three stages: a convolutional layer, a pooling layer, and a fully-connected (FC) layer.

The convolutional layer extracts meaningful features through convolution, and the pooling layer is a process of subsampling to reduce the generated feature values. The fully connected layer is classified using the NN completely connected to the output layer after transforming the feature set from the convolutional layer and the pooling layer into one-dimensional data.

CNN technology has been applied to image classification/detection problems using images or video data. Even for time series data such as acoustics, there are two approaches to applying CNN models. First, as a method of transforming one-dimensional data into two-dimensional data and using it, for example, there is a method of converting one-dimensional sound data into a spectrogram image and then inputting it into a CNN. Another method is to apply one-dimensional data to one-dimensional convolution rather than a two-dimensional convolution process.

The present invention is a technique for performing seismic classification or seismic detection based on a state-based convolutional neural network.

The attention-based convolutional neural network-based seismic classification technique presented in the present invention includes a preprocessing step, a feature extraction step (Conv 1 ~ Conv 8), and a classification step (Fully Connect 1, Fully Connect 2) as shown in FIG. do. Each of the steps is processed by the preprocessing unit 110 , the feature extracting unit 130 , and the classifying unit 150 .

The pre-processing unit 110 performs data pre-processing of Equation 1 below and a centering process when cases in which seismic data ranges are different due to the characteristics of each station. In general, the centering process can achieve the effect of geometrically moving the data cluster to the origin for all dimensions.

[Equation 1]

Here, xm is the observed data, ym is the center-shifted data, and M is the total number of samples taken from one data.

The feature extraction unit 130 extracts a feature map by non-linearly transforming the pre-processed seismic data, and the non-linearly transformed seismic data based on the attention technique modeling the interdependence between channels of the feature map. Three or more values are measured, and the measured importance value is corrected by element-by-element multiplication with the learned feature map, and down-sampling through max-pooling based on the feature value. It includes a plurality of convolutional layers.

The attention-based CNN structure proposed by the present invention is shown in FIGS. 2 to 4 . FIG. 3 is a block diagram illustrating first and eighth convolutional layers in the seismic event classification apparatus using the attention-based neural network of FIG. 1 . FIG. 4 is a block diagram illustrating second to seventh convolutional layers in the seismic event classification apparatus using the attention-based neural network of FIG. 1 . 5 is a block diagram illustrating an attention module of the earthquake event classification apparatus using the attention-based neural network of FIG. 1 .

Referring to FIG. 3 , the model input uses three-channel seismic data and uses input data of 1000 samples per channel. 1000-sample input data represents data of a 10-second interval of 100 Hz sampling. The proposed model consists of eight convolutional layers, two fully-connected (FC) layers, and a softmax function.

In the eight convolutional layers, 1D convolution is performed as shown in Equation 2 below, and the convolution result is subjected to non-linear mapping through a Rectified Linear Unit (ReLU) function as shown in Equation 3 below. The filter size of 1D convolution uses a 3x1 shape, and 32 channels can be applied for each layer.

[Equation 2]

Here, F ^l _c,t denotes a feature map in 1 layer, and c and t denote a channel and a sample time, respectively. Also, W _l refers to a convolution filter in layer l.

[Equation 3]

Referring to FIGS. 3 and 4 , each convolutional layer of the feature extraction unit includes a nonlinear transformation unit 133 that nonlinearly transforms the preprocessed earthquake data, a caution module 135, and a max pooling (max) based on the feature value. -pooling) includes a max pooling unit 137 for down-sampling (down-sampling).

The attention module 135 measures the importance of a learned feature based on an attention technique that models the interdependence between channels of the feature map using nonlinearly transformed seismic data, and outputs a feature value.

The first and last convolutional layers, the first and eighth convolutional layers, further include a batch normalization unit 131 before the nonlinear transformation unit 133 as shown in FIG. 3 .

The nonlinear mapping result of the nonlinear transformation unit 133 measures the importance of the learned feature based on the attention technique shown in FIG. 5 . The measured importance yields an improved feature value through element-wise multiply with the learned feature.

The calculated feature value is subjected to a down-sampling process through max-pooling. The interval of max pooling is set to, for example, 2, and each time this process is performed, the feature information can be reduced by 1/2.

The attention module 135 is a means for enhancing the feature by extracting the main components from the feature. In the present invention, as shown in FIG. 5 , an improved attention module based on SENET (Squeeze and Excitation Network) is proposed.

SENET improved the expressive ability of the network by modeling the interdependence between channels of the feature map. SENET consists of a squeeze operation unit 310, excitation operation units 510 to 570, and a scale unit 710 that corrects the feature value through element-by-element multiplication of the measured importance value with the learned feature map. do.

The squeeze operation unit 310 compresses the feature map while maintaining global information by using global average pooling in each channel. The activation calculation units 510 to 570 are used by the two fully-connected (FC) layers 510 and 550 and the sigmoid activation function 570 to adaptively recalibrate the attention value according to the importance of each channel. .

The first fully connected layer 510 represents a dimensionality reduction layer having a reduction rate r, and may further include a Rectified Linear Unit (ReLU) 530 . The second fully connected layer 550 represents a dimensional reconstruction layer that restores the channel dimension of the feature map.

The final attention value adjusts the importance of the channel to a value between 0 and 1 through the sigmoid activation function 570 . In the existing SENET, a squeeze operation was used using global average pooling as shown in Equation 4 below.

In the present invention, a global weighted average pooling method using a contrast stretching method used for image enhancement is adopted as shown in Equation 5 below. Contrast stretching is a method of improving the visibility of an image by spreading the histogram of a low-contrast image distribution to achieve a wide range of contrast values. When this concept is applied to the feature map, it is possible to include more distinct channel-specific statistical information.

[Equation 4]

[Equation 5]

이며,

과

는 특징맵에서 채널 c의 최대값과 최소값을 나타낸다. Here, F _c (i) represents the feature map of channel c, and the length of each channel is L. In addition,

is,

class

denotes the maximum and minimum values of channel c in the feature map.

Hereinafter, a batch normalization unit 131 included in the first and last convolutional layers will be described.

As the network layer increases, gradient vanishing and exploding problems are major issues in deep networks. Various attempts have been made to solve this problem through changes in the activation function, initialization, and learning rate.

Batch normalization takes an approach that solves the problem of gradient vanishing/gradient exploding in the essential training process. Batch normalization was devised by considering the whitening method that normalizes the input distribution of each layer in the form of mean 0 and standard deviation 1 in order to prevent the internal covariance shift phenomenon in which the input distribution is different for each layer.

After finding the mean and variance for each feature, normalization is performed, and a transformation process of adding a scale factor and a shift factor to the normalized values is performed. The scale factor and shift factor parameters can be derived in the training process in mini-batch units.

In the present invention, drop-out regularization is applied to the first fully-connected (FC) layer in order to prevent overfitting of the classifier 150 .

Dropout activates each neuron according to a probability value and applies it to learning. In the present invention, the network can be trained by setting p = 0.5 in the first fully accessed layer.

The present invention robustly extends the existing CNN-based deep learning technique to noise and overfitting through attention module, batch normalization, and dropout regularization.

The present invention can be utilized for effective earthquake warning of the national earthquake comprehensive information system. Specifically, by building a robust earthquake detection system in a situation where noise and artificial earthquakes are mixed, it is possible to contribute to the protection of people's lives and safety by applying accurate and low-error earthquake detection information and promptly and accurately alerting and responding to them. do. In addition, by distinguishing from artificial earthquakes, appropriate responses can be made according to earthquake characteristics and circumstances, and the possibility of false alarms can be reduced.

In addition, the present invention is utilized to obtain detailed seismic data information. When the seismic detector developed through the present invention is applied to an earthquake observation system, it is possible to collect high-accuracy seismic data with a low probability of false detection due to noise and artificial earthquakes. This data can be provided as applicable data for seismic observations and other seismic studies.

6 is a flowchart of a seismic event classification method using an attention-based neural network according to an embodiment of the present invention.

The seismic event classification method using the attention-based neural network according to the present embodiment may proceed in substantially the same configuration as the apparatus 10 of FIG. 1 . Accordingly, the same components as those of the device 10 of FIG. 1 are given the same reference numerals, and repeated descriptions are omitted.

In addition, the seismic event classification method using the attention-based neural network according to the present embodiment may be executed by software (application) for performing seismic event classification using the attention-based neural network.

Referring to FIG. 6 , in the seismic event classification method using the attention-based neural network according to the present embodiment, input seismic data is preprocessed by centering it (step S100 ).

In step S100, when the seismic data ranges are different due to the characteristics of each station, Equation 1 and the data preprocessing process of the centering process are performed. In general, the centering process can achieve the effect of geometrically moving the data cluster to the origin for all dimensions.

Next, as a feature extraction step, the preprocessed seismic data is non-linearly transformed through a plurality of three or more convolutional layers to extract a feature map (step S310).

For example, it can be performed through eight convolutional layers, each convolutional layer performs 1D convolution as shown in Equation 2 above, and the convolution result is nonlinear mapping ( mapping). The filter size of 1D convolution uses a 3x1 form, and 32 channels can be applied for each layer.

The importance of the learned feature is measured based on the attention technique that models the interdependence between channels of the feature map using the nonlinearly transformed seismic data (step S330).

In step S330, the importance of each channel by using a squeeze operation that compresses the feature map while maintaining global information in each channel, two fully-connected (FC) layers, and a sigmoid activation function An excitation operation that adaptively recalibrates the importance value is performed.

The squeeze operation uses a global weight average pooling method that spreads a histogram of a low-contrast image distribution based on contrast stretching.

The excitation operation may include passing the compressed feature map data through a dimensionality reduction layer having a preset reduction rate and restoring the data passing through the dimension reduction layer to the channel dimension of the feature map. have.

The step of non-linearly mapping the input data of the pre-processed seismic data through a plurality of convolutional layers of three or more includes placing the pre-processed seismic data in the first convolutional layer and the last convolutional layer among the plurality of convolutional layers. The normalization step can be rougher.

At this time, the average and variance are obtained for each feature value, then normalized, and the normalized feature values are converted by adding a scale factor and a shift factor.

The measured importance value corrects the feature value through element-by-element multiplication with the learned feature map (step S350), and down-sampling through max-pooling based on the feature value (step S350) (step S350). S370). For example, the interval of max pooling is set to, for example, 2, and each time this process is performed, the feature information can be reduced by 1/2.

An earthquake event is classified by normalizing the down-sampled feature value (step S500).

In step S500 , drop-out regularization may be applied to the first fully-connected (FC) layer in order to prevent overfitting.

Dropout activates each neuron according to a probability value and applies it to learning. In the present invention, the network can be trained by setting p = 0.5 in the first fully accessed layer.

Hereinafter, experimental results for verifying the effects of the present invention will be described. For the experiment, seismic data and noise data were collected using seismic data information that occurred in Korea from January 1, 2016 to July 31, 2018. An earthquake event and noise database was constructed based on the 24-hour seismic observation data and seismic event list file provided by the Korea Meteorological Administration.

Only 100 sample data were used for earthquake data, and the database consisted of strong earthquakes (magnitude 2.0 or greater), micro-earthquakes (less than 2.0 magnitudes), artificial earthquakes, and noise. Data from 2016 to 2017 were used as a training dataset, and data from 2018 was used as a test dataset.

The seismic event extraction method extracts the data for 10 seconds from the time the earthquake was observed at each station. However, in order to increase the amount of data collected, in the present invention, data is collected for a total of 13 seconds from 3 seconds before the earthquake, and a method of sliding the extraction window at intervals of 1 second is applied.

The collection of noise events was applied in two ways. For the date recorded in the earthquake event list, noise data was extracted by random sampling method except for 1 hour before and after the occurrence time, and for the date not recorded in the earthquake event list, noise data is randomly selected within 24 hours. extracted. Table 1 shows the composition of the event data set used in the simulation.

event (One)
(Learn/Test) (2)
(Learn/Test) (3)
(Learn/Test) Earthquake(1) vs Noise(2) 21,340/3,248 22,523/5,752 Earthquake(1) vs Noise(2) 10,772/688 22,523/2,312 Smile(1) vs Noise(2) 8,328/1,932 22,523/4,068 Artificial(1) vs Noise(2) 2,240/628 22,523/2,372 Natural(1)vs Artificial(2)vs Noise(3) 19,100/2,620 2,240/628 22,523/5,752

This simulation was performed based on NVIDIA Geforce GTX 1080Ti GPU and Tensorflow, and a total of 300 epochs of network training were performed. When training the model, the ADAM method was used for optimization, and the learning rate was given as 5×10 ^-3 . To analyze the performance of the invention technique, a comparative analysis with ConvNetQuake was performed, and the accuracy, true positive rate, and false positive rate indices were used for performance measurement as shown in Equations 6 to 8.

[Equation 6]

[Equation 7]

[Equation 8]

Here, TP represents the number of true positives, TN represents the number of true negatives, and FP represents the number of false positives.

Table 2 below compares the performance of the method proposed in the present invention (unit %).

ConvNetQuake Suggested method 1
(via GAP) Suggested method 2
(via GWAP) acc TPR FPR acc TPR FPR acc TPR FPR earthquake vs noise 93.2 89.2 4.5 97.0 95.0 1.8 97.6 95.7 1.3 earthquake vs noise 95.8 89 2.1 98.2 96.4 1.3 98.8 97.4 0.8 smile vs noise 93.9 84.5 1.6 98.2 95.7 0.6 98.2 96.3 0.9 artificial vs noise 93.7 74.5 1.2 96.1 83.0 0.4 96.7 84.9 0.2 Nature vs Artificial vs Noise 89.1 84.6/
45.2/
95.9 6.5/
1.2/
14.2 95.3 93.5/
72.9/
98.5 2.6/
1.1/
5.2 95.6 94.6/74.2/
98.5 2.7/
1.0/
4.2

Table 2 In the experiment, the baseline structure was compared to the existing ConvNetQuake. The method proposed in Table 2 showed better performance than the baseline method for all event classifications.

Proposed method 2 improved the accuracy by an average of 4.24% compared to the existing baseline. Also, in case of binary classification, TPR improvement was remarkable, and in case of multiple classification, TPR improvement and FPR decrease were shown. The artificial seismic event did not show as good performance as other event classifications, and it is estimated that the data quantity and data length are insufficient for this part.

In addition, in the case of the attention module according to the squeeze operation, the method applying the proposed GWAP showed better performance in all event classification than the method applying the existing GAP, and showed an average 0.42% accuracy improvement.

Such a seismic event classification method using an attention-based neural network may be implemented as an application or implemented in the form of program instructions that may be executed through various computer components and recorded in a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc. alone or in combination.

The program instructions recorded on the computer-readable recording medium are specially designed and configured for the present invention, and may be known and available to those skilled in the computer software field.

Examples of the computer-readable recording medium include hard disks, magnetic media such as floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floppy disks. media), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like.

Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform processing according to the present invention, and vice versa.

Although the above has been described with reference to the embodiments, those skilled in the art can variously modify and change the present invention within the scope without departing from the spirit and scope of the present invention described in the claims below You will understand.

The present invention can be utilized in the disaster prevention IT market. Disaster situations related to earthquakes are increasing worldwide. Korea has also been experiencing the Gyeongju and Pohang earthquakes, and much interest and investment is being made in disaster prevention systems.

In addition, the present invention is expected as a technology that can be widely operated in countries (meteorological offices of each country) or allied organizations rather than companies that generate profits as disaster prevention-related technologies.

10: Seismic event classification device using attention-based neural network
110: preprocessor
130: feature extraction unit
150: classification unit
131: batch normalization unit
133: non-linear transformation unit
135: attention module
137: Max pulling unit
310: squeeze (Squeeze) operation unit
510: first fully connected layer
530: ReLU
550: second fully connected layer
570: sigmoid activation function
710: scale unit

Claims (15)

pre-processing by centering the input earthquake data;
extracting a feature map by non-linearly transforming the preprocessed seismic data through a plurality of convolutional layers of three or more;
Measuring the importance of the learned feature based on the attention technique that models the interdependence between channels of the feature map on the nonlinearly transformed seismic data;
correcting the feature value through element-by-element multiplication of the measured importance value with the learned feature map;
down-sampling through max-pooling based on the feature value; and
Classifying seismic events by normalizing the down-sampled feature values; Containing, a seismic event classification method using a caution-based neural network.
The method of claim 1, wherein measuring the importance of the learned feature comprises:
A seismic event classification method using a caution-based neural network, comprising the step of performing a squeeze operation that compresses a feature map while maintaining global information in each channel.
The method of claim 2, wherein performing the squeeze operation comprises:
A method of classifying seismic events using attention-based neural networks using a global weight average pooling method that spreads a histogram of a distribution of low-contrast images based on contrast stretching.
The method of claim 2, wherein measuring the importance of the learned feature comprises:
Using two fully-connected (FC) layers and a sigmoid activation function to adaptively recalibrate the importance value according to the importance of each channel, the method further comprising the step of calculating an excitation operation , A seismic event classification method using attention-based neural networks.
The method of claim 4, wherein the excitation calculation comprises:
passing the compressed feature map through a dimensionality reduction layer having a preset reduction rate; and
and restoring the data that has passed through the dimension reduction layer to the channel dimension of the feature map.
The method of claim 1, wherein the nonlinear mapping of the input data of the preprocessed seismic data through three or more convolutional layers comprises:
Seismic event classification method using an attention-based neural network, further comprising the step of batch-normalizing the preprocessed seismic data in a first convolutional layer and a last convolutional layer among the plurality of convolutional layers.
The method of claim 6, wherein the batch normalizing the feature value comprises:
A seismic event classification method using an attention-based neural network that calculates the mean and variance for each feature value, then normalizes it and converts the normalized feature values by adding a scale factor and a shift factor.
The method of claim 1, wherein classifying the seismic event by normalizing the down-sampled feature value comprises:
An attention-based neural that passes through first and second fully-connected (FC) layers, and the first fully-connected layer performs a dropout normalization process that activates according to the probability value of each neuron and applies it to learning. A method of classifying earthquake events using networks.
9. The method according to any one of claims 1 to 8,
A computer-readable storage medium in which a computer program for performing the seismic event classification method using the attention-based neural network is recorded.
a pre-processing unit for pre-processing the input earthquake data by centering it;
The preprocessed seismic data is non-linearly transformed to extract a feature map, and the non-linearly transformed seismic data is used to measure the importance of the learned feature based on the attention technique that models the interdependence between channels of the feature map, and the measured importance value is A feature comprising three or more convolutional layers that correct a feature value through element-by-element multiplication with the learned feature map, and down-sample through max-pooling based on the feature value extraction unit; and
A classification unit including first and second fully-connected (FC) layers for classifying seismic events by normalizing the down-sampled feature values.
The method of claim 10, wherein each convolutional layer of the feature extraction unit comprises:
a non-linear conversion unit for non-linearly converting the pre-processed earthquake data;
an attention module that measures the importance of features learned based on an attention technique that models the interdependence between channels of the feature map using nonlinearly transformed seismic data and outputs a feature value; and
A seismic event classification device using a state-based neural network, including; a max pooling unit for down-sampling through max-pooling based on the feature value.
11. The method of claim 10, wherein the first convolutional layer and the last convolutional layer of the feature extraction unit,
Attention-based neural network, which further includes a batch normalizer that calculates the mean and variance for each feature value, then undergoes normalization, and converts the normalized feature values by adding a scale factor and a shift factor seismic event classification device.
12. The method of claim 11, wherein the attention module,
a squeeze operation unit that compresses the nonlinearly transformed seismic data while maintaining global information in each channel;
an excitation operation unit that adaptively re-corrects the importance value according to the importance of each channel using two fully-connected (FC) layers and a sigmoid activation function; and
The seismic event classification device using an attention-based neural network, including; a scale unit that corrects the feature value through element-by-element multiplication of the measured importance value with the learned feature map.
14. The method of claim 13, wherein the exit calculation unit,
a dimensionality reduction unit passing the compressed feature map through a dimensionality reduction layer having a preset reduction rate; and
An earthquake event classification device using an attention-based neural network, comprising a dimension restoration unit that restores the data that has passed through the dimension reduction layer to the channel dimension of the feature map.
The method of claim 10, wherein the first fully connected layer of the classification unit,
An earthquake event classification device using an attention-based neural network that performs a dropout regularization process that activates according to the probability value of each neuron and applies it to learning.

Images