KR102635608B1 - Earthquake Event Classification Method and Apparatus Using Multi-task Deep Learning - Google Patents

Abstract

The present invention discloses a method and device for classifying earthquake events using multi-task deep learning. According to the present invention, a preprocessing unit that centrally preprocesses earthquake data; a feature extraction unit that receives the pre-processed seismic data and extracts a feature map with multiple channels through 1D convolution operation in multiple convolution layers and non-linear mapping using a ReLU function; A two-class classifier that classifies whether or not there is an earthquake through the extracted feature map; And a three-class classifier that classifies the type of earthquake through the extracted feature map, wherein each of the plurality of convolutional layers includes an attention module that measures the importance of each channel of the feature map extracted through the non-linear mapping ( An earthquake event classification device is provided, which includes an attention module) and extracts an improved feature map by multiplying the measured importance and the extracted feature map for each element.

Description

Earthquake Event Classification Method and Apparatus Using Multi-task Deep Learning}

The present invention relates to a method and device for classifying earthquake events using multi-task deep learning.

Convolutional Neural Network (CNN) is a type of neural network and has a chain structure of linear combination and nonlinear transformation through convolution operations.

In the case of multi-layer perceptron (MLP), when the dimension of the learning data (e.g., image input) increases, the amount of parameters to be learned increases exponentially, and a large amount of data is required to learn the parameters.

In contrast, CNN can significantly reduce the number of parameters to be learned through convolution operations with the characteristics of local connection and weight sharing, and can automatically generate convolution filters that meet the purpose through learning.

Generally, CNN consists of three stages: convolutional layer, pooling layer, and fully-connected (FC) layer.

Here, the convolution layer extracts meaningful features (feature maps) through convolution operations and nonlinear transformation, the pooling layer performs a subsampling process to reduce the generated features, and the FC layer extracts the feature set. After transforming into one-dimensional data, classification is performed using a neural network fully connected to the output layer.

Meanwhile, the existing earthquake classification algorithm used STA (Short Time Average) and LTA (Long Time Average) methods, which are representative methods for detecting earthquakes through rate thresholds, but are unsuitable in low SNR environments.

In addition, the autocorrelation-based method detects earthquakes based on the correlation between seismic waveforms repeated in a single area. Although it has excellent performance, it has limitations in usability due to computational overload and long-term signals.

The template matching method is a technique proposed to improve the computational overload problem of the autocorrelation method. After setting an earthquake template, the earthquake is detected through correlation between the template and the input seismic 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, the FAST (Fingerprint And Similarity Thresholding) method was 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.

Recently, a variety of earthquake detection techniques using deep learning have been reported.

The problem of classifying an earthquake event involves not only classifying whether there is an earthquake, but also classifying the type of earthquake. In general, compared to large-scale earthquakes, the detection of small-scale micro-earthquakes has unsatisfactory classification performance with noise due to the low SNR environment, and artificial earthquakes such as artificial blasting and nuclear testing may have poor classification performance with natural micro-earthquakes. there is.

Republic of Korea Patent Publication 10-2020-0061307

In order to solve the problems of the prior art described above, the present invention proposes a seismic event classification method and device using multi-task deep learning that can effectively classify seismic data and noise and types of earthquakes.

In order to achieve the above-mentioned object, according to an embodiment of the present invention, there is provided an earthquake event classification device, comprising: a pre-processing unit for centering and pre-processing seismic data; a feature extraction unit that receives the preprocessed seismic data and extracts a feature map with multiple channels through 1D convolution operation in multiple convolution layers and nonlinear mapping using a ReLU function; A two-class classifier that classifies whether or not there is an earthquake through the extracted feature map; And a three-class classifier that classifies the type of earthquake through the extracted feature map, wherein each of the plurality of convolutional layers includes an attention module that measures the importance of each channel of the feature map extracted through the non-linear mapping ( An earthquake event classification device is provided that includes an attention module) and extracts an improved feature map by multiplying the measured importance and the extracted feature map for each element.

Among the plurality of convolutional layers, the first convolutional layer into which the preprocessed seismic data is input may perform batch normalization between the 1D convolution operation and nonlinear mapping through the ReLU function.

The 2-class classifier and 3-class classifier may be composed of two fully connected layers and a softmax function.

The three-class classifier converts the feature maps output by at least a portion of the plurality of convolutional layers into feature vectors by vectorizing them, and injects the converted feature vectors into inputs of the two fully connected layers to determine the type of earthquake. can be classified.

The attention module can measure the importance based on SENET (Squeeze and Excitation Network).

The attention module can compress the feature map while maintaining global information through a global weight average pooling method using a contrast stretching method.

According to another aspect of the invention, a method for classifying a seismic event in a device comprising a processor and memory, comprising:

centering preprocessing seismic data; Receiving the pre-processed seismic data and extracting a feature map with a plurality of channels through 1D convolution operation in a plurality of convolution layers and non-linear mapping using a ReLU function; Classifying whether or not there is an earthquake using the extracted feature map using a two-class classifier; And classifying the type of earthquake through the extracted feature map through a three-class classifier, wherein each of the plurality of convolutional layers measures the importance of each channel of the feature map extracted through the non-linear mapping. An earthquake event classification method is provided that includes an attention module and extracts an improved feature map by multiplying the measured importance and the extracted feature map for each element.

According to the present invention, key features suitable for the presence and type of earthquake can be extracted through multi-task learning, and earthquake detection even in a low SNR environment through feature extraction with improved discriminability for each class through attention-based feature extraction. It has the advantage of improving performance.

In addition, according to the present invention, by providing accurate and rapid earthquake information, it is possible to reduce anxiety about earthquakes in normal times and improve the reliability of warnings, and to encourage construction in safe locations when planning the construction of major industrial facilities based on earthquake detection analysis information. Alternatively, the possibility of a large-scale disaster can be reduced through preparations such as earthquake-resistant design.

Furthermore, by analyzing accumulated micro-earthquake detection information, it is possible to derive areas of concern, and by intensively monitoring nearby areas, quick warning and response to earthquake damage is possible.

Figure 1 is a diagram illustrating the structure of a multi-task-based earthquake event classification network according to an embodiment of the present invention.
Figure 2 is a diagram showing a convolutional layer to which an attention module according to this embodiment is applied.
Figure 3 is a diagram showing the configuration of a SENET-based attention module.
Figure 4 is a diagram illustrating an attention module-based feature merging structure according to this embodiment.

Since the present invention can make various changes and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail.

However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all changes, equivalents, and substitutes included in the spirit and technical scope of the present invention.

This embodiment presents a method for classifying whether there is an earthquake and the type of earthquake based on multi-task deep learning.

Figure 1 is a diagram illustrating the structure of a multi-task-based earthquake event classification network according to an embodiment of the present invention.

As shown in FIG. 1, the device according to this embodiment may include a preprocessor 100, a feature extraction unit 102, a 2-class classifier 104, and a 3-class classifier 106.

A device according to this embodiment may be a device including a processor and memory.

Here, the processor may include a central processing unit (CPU) capable of executing a computer program or another virtual machine.

Memory may include non-volatile storage devices, such as non-removable hard drives or removable storage devices. Removable storage devices may include compact flash units, USB memory sticks, etc. Memory may also include volatile memory, such as various types of random access memory.

Program instructions executable by the processor 100 are stored in this memory.

In the case of seismic data, the range of seismic data may vary due to the characteristics of each station.

Considering this, the preprocessing unit 100 performs earthquake data preprocessing using Equation 1 below and the centering process.

Through the centering process, the effect of geometrically moving a cluster of data to the origin in all dimensions can be achieved.

Here, M is the total number of samples acquired from one data.

Input according to this embodiment uses 3-channel seismic data and uses input data of 1000 samples per channel. 1000 sample input data represents data in a 10 second interval of 100Hz sampling.

The feature extraction unit 102 receives preprocessed seismic data and uses a plurality of convolutional layers (Conv1 to Conv8) to extract feature maps with multiple channels for each filter through non-linear mapping using 1D convolution operation and the ReLU function. Includes.

Figure 1 exemplarily shows that the feature extraction unit 102 includes eight convolution layers.

Additionally, the network according to this embodiment may include a 2-class classifier 104 that classifies whether there is an earthquake through the extracted feature map and a 3-class classifier 106 that classifies whether there is an earthquake through the extracted feature map.

As shown in FIG. 1, each classifier 104 and 106 consists of two fully-connected (FC) layers and a softmax function.

The eight convolution layers (Conv1 to Conv8) perform a 1D convolution operation as shown in Equation 2 below, and the results of the convolution operation undergo non-linear mapping through the ReLU (Rectified Linear Unit) function as shown in Equation 3.

In the first convolution layer (Conv 1), a batch normalization process is performed between the convolution operation and ReLU, as shown in FIG. 2. The filter size of 1D convolution operation uses a 3x1 format, and a channel size of 32 can be applied to each layer.

은 컨볼루션 레이어

의 특징맵을 나타내고,

은 입력 파형을 나타내고,

은 차원

을 갖는 레이어

의 필터 가중치를 나타내고,

은 레이어

의 채널 수이다. here,

is a convolutional layer

represents the feature map of

represents the input waveform,

silver dimension

layer with

represents the filter weight of

silver layer

is the number of channels.

For each channel of the non-linearly mapped feature map, importance is measured based on the attention module as shown in FIG. 3.

As shown in Figure 3, the importance of each channel is calculated by element-wise multiplying with the feature map to produce an improved feature map and down-sampling through max-pooling. It goes through a sampling process. The max pooling interval is set to 2, and each time this process is performed, the feature information is reduced by 1/2. In this embodiment, drop-out regularization is applied to the first fully connected layer (FC 1) to prevent overfitting.

The feature extraction unit 102 extracts shared features as a hidden layer, the 2-class classifier 104 classifies earthquakes and noise through the shared features, and the 3-class classifier 106 classifies small earthquakes, strong earthquakes, and noise.

In the two-class classification process according to this embodiment, the classification task is performed using only the feature information of the last convolutional layer.

In more detail, the two-class classifier 104 converts the feature map of the last convolutional layer (Conv 8) into a feature vector through a vectorization process, and injects the converted feature vector as the input of two fully connected layers. Two-class classification is performed using the softmax function.

The 3-class classifier 106 uses the fact that high-frequency features are mainly extracted from the initial convolution layer, and as the layer becomes deeper, low-frequency features are mainly extracted, and the 6th to 8th convolution layers (Conv 6 to Conv 8) ) to form a feature vector by merging the feature maps of the latter convolutional layers.

That is, as shown in FIG. 4, instead of simply connecting the feature maps of each layer, a method of merging the feature maps of multiple convolutional layers using an attention module is used.

The attention-based feature merging method according to this embodiment constructs a feature vector that focuses more on important parts of the task.

와 3 클래스 분류기(106)의 손실

의 합으로 구성되며, 네트워크 가중치 W의 L2 norm을 통한 정규화를 적용한다. The loss function of the multi-task learning structure according to this embodiment is the loss of the two-class classifier 104 as shown in Equation 4 below:

and the loss of the 3-class classifier 106.

It is composed of the sum of , and normalization through the L2 norm of the network weight W is applied.

The attention module is a means of improving features by extracting key components from the features.

In this embodiment, as shown in FIG. 4, an improved attention module based on SENET (Squeeze and Excitation Network) is applied.

SENET consists of a squeeze operation and an excitation operation.

In general, the Squeeze operation uses global average pooling (GAP) in each channel to compress the feature map while maintaining global information. Excitation operation uses two fully connected layers and a sigmoid activation function to adaptively recalibrate the attention value according to the importance of each channel.

The first fully connected layer (FC 1) represents a dimension reduction layer with a reduction rate r, and the second fully connected layer (FC 2) represents a dimension restoration layer that restores the channel dimension of the feature map.

The final attention value adjusts the importance (weight) of the channel to a value between 0 and 1 through a sigmoid activation function.

While the existing SENET uses a squeeze operation using GAP, in this embodiment, Global Weight Average Pooling using the contrast stretching method used for image enhancement as shown in Equation 5 above is used. ) suggests a method.

Contrast stretching is a method of improving the visibility of an image by spreading the histogram of the low-contrast image distribution to achieve a wide range of contrast values. This makes it possible to perform a squeeze operation that includes clear statistical information for each channel.

,

와

는 특징맵에서 채널 c의 글로벌 최대값과 최소값이다. Here, s is the squeeze operation vector, L is the length,

,

and

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

Below, the experimental process for evaluating the performance of this embodiment will be described.

For the experiment, seismic data and noise data were collected using earthquake data information that occurred in Korea from January 1, 2016 to July 31, 2018.

A seismic event and noise database was constructed based on 24-hour seismic observation data and seismic event list files provided by the Korea Meteorological Administration.

Earthquake data used only 100 sample data, and the database was divided into medium-scale earthquakes (magnitude 3.0 or higher), micro-earthquakes (magnitude less than 3.0), and noise. However, in the case of domestic earthquake data, because the mid-scale earthquake data is very small, the mid-scale earthquake data was additionally used as earthquake data that occurred in the Pacific Rim between March 2019 and September 2019 provided by IRIS.

Extraction of earthquake events extracts data from a 10-second section from the time the earthquake was observed for each observatory. However, in order to increase the amount of collected data, this experiment collected data for a total of 13 seconds starting 3 seconds before the earthquake occurred, and applied a method of sliding the extraction window at 1-second intervals.

Collection of noise events was applied in two ways. For dates recorded in the earthquake event list, 10 seconds of noise data was extracted by random sampling from the portion excluding the 1 hour before and after the occurrence time. For dates not recorded in the earthquake event list, noise data was randomly sampled within 24 hours of data. was extracted. The composition of the event dataset used in the simulation experiment is shown in Table 1.

By event [One] [2] [3] Earthquake[1] vs Noise[2] 25432/4720 40000/20000 Small scale [1] vs Mesoscale [2] vs Noisy [3] 19604/2740 5288/1980 40000/20000

는 10^-5을 주었다. This simulation experiment was performed based on NVIDIA Geforce GTX 1080Ti GPU and Tensorflow, and network training was performed for a total of 300 epochs. When training the model, the ADAM method was used for optimization, the learning rate was 10 ^-3 , and the L2 regularization parameter was

gave 10 ^-5 .

For performance analysis, a comparative analysis was performed with the ConvNetQuake-based single task structure, and the accuracy, true positive rate, and false positive rate indices were used to measure performance as shown in Equation 6.

The performance comparison results are shown in Table 2.

event
class Single-tasking structure Multi-tasking structure Accuracy TPR FPR Accuracy TPR FPR earthquake vs
Noise 98.6 96.1 0.9 99.1 96.7 0.4 Smile vs Mesoscale vs Noise 98.2 91.9/
94.3/
99.4 0.5/
0.6/
4.8 98.7 95.4/
93.9/
99.7 0.4/
0.4/
3.3

The baseline structure was compared and analyzed with the existing single task structure based on ConvNetQuake. In Table 2, the method according to this example showed better performance than the baseline method for all event classifications. The proposed method showed improved performance over the existing baseline, and in particular, the performance in terms of TPR was significantly improved in the 3-class classification results.

Additionally, as shown in Table 3, the multi-task learning structure based on the proposed attention-based feature merging showed better performance than the simple feature merging method.

event
class Simple MAX-based FA Attention-based FA Accuracy TPR FPR Accuracy TPR FPR earthquake vs
Noise 98.9 96.6 0.6 99.1 96.7 0.4 Smile vs Mesoscale vs Noise 98.5 94.6/
93.0/
99.5 0.5/
0.6/
3.9 98.7 95.4/
93.9/
99.7 0.4/
0.4/
3.3

The above-described embodiments of the present invention have been disclosed for illustrative purposes, and those skilled in the art will be able to make various modifications, changes, and additions within the spirit and scope of the present invention, and such modifications, changes, and additions will be possible. should be regarded as falling within the scope of the patent claims below.

Claims (7)

A seismic event classification device, comprising:
A preprocessing unit that centrally preprocesses seismic data;
a feature extraction unit that receives the pre-processed seismic data and extracts a feature map with multiple channels through 1D convolution operation in multiple convolution layers and non-linear mapping using a ReLU function;
A two-class classifier that classifies whether or not there is an earthquake through the extracted feature map; and
It includes a 3-class classifier that classifies the type of earthquake through the extracted feature map,
Each of the plurality of convolutional layers includes an attention module that measures the importance of each channel of the feature map extracted through the non-linear mapping, and multiplies the measured importance and the extracted feature map for each element. Extract the improved feature map,
The two-class classifier classifies whether or not there is an earthquake through the feature map output by the last convolutional layer among the plurality of convolutional layers,
The three-class classifier converts feature maps output by at least some of the plurality of convolutional layers located at the rear end into feature vectors by vectorizing them, and injects the converted feature vectors into inputs of two fully connected layers to perform earthquake analysis. Classify the types of
The attention module measures the importance based on SENET (Squeeze and Excitation Network),
The attention module is a seismic event classification device that compresses the feature map while maintaining global information through a global weight average pooling method using a contrast stretching method. According to paragraph 1,
Among the plurality of convolutional layers, the first convolutional layer into which the preprocessed seismic data is input performs batch normalization between the 1D convolution operation and nonlinear mapping through the ReLU function. According to paragraph 1,
The 2-class classifier and 3-class classifier are earthquake event classification devices consisting of two fully connected layers and a softmax function. delete delete delete 1. A method for classifying seismic events in a device comprising a processor and memory, comprising:
centering preprocessing seismic data;
Receiving the pre-processed seismic data and extracting a feature map with a plurality of channels through 1D convolution operation in a plurality of convolution layers and non-linear mapping using a ReLU function;
Classifying whether or not there is an earthquake using the extracted feature map using a two-class classifier; and
3 Including the step of classifying the type of earthquake through the extracted feature map through a class classifier,
Each of the plurality of convolutional layers includes an attention module that measures the importance of each channel of the feature map extracted through the non-linear mapping, and multiplies the measured importance and the extracted feature map for each element. Extract the improved feature map,
The two-class classifier classifies whether or not there is an earthquake through the feature map output by the last convolutional layer among the plurality of convolutional layers,
The three-class classifier converts feature maps output by at least some of the plurality of convolutional layers located at the rear end into feature vectors by vectorizing them, and injects the converted feature vectors into inputs of two fully connected layers to perform earthquake analysis. Classify the types of
The attention module measures the importance based on SENET (Squeeze and Excitation Network),
The attention module is a seismic event classification method that compresses the feature map while maintaining global information through a global weight average pooling method using a contrast stretching method.