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

1. An earthquake event classification method using an attention-based neural network, the method comprising:

preprocessing input earthquake data by centering;

extracting a feature map by nonlinearly converting the preprocessed earthquake data through a plurality of convolution layers having three or more layers;

measuring an importance value of a learned feature of the nonlinear-converted earthquake data based on an attention technique in which interdependence between channels of the feature map is modeled;

correcting a feature value of the measured importance value through element-wise multiply with the learned feature map;

performing down-sampling through max-pooling based on the corrected feature value; and

classifying an earthquake event by regularizing the down-sampled feature value.

2. The earthquake event classification method using an attention-based neural network according to claim 1 ,

wherein said measuring of the importance value of the learned feature includes:

performing a squeeze operation to compress the feature map while maintaining global information at each channel.

3. The earthquake event classification method using an attention-based neural network according to claim 2 ,

wherein said performing of the squeeze operation uses a global weight average pooling method that diffuses a histogram with a low contrast image distribution based on contrast stretching.

4. The earthquake event classification method using an attention-based neural network according to claim 2 ,

wherein said measuring of the importance value of the learned feature further includes:

performing an excitation operation to adaptively recalibrate the importance value according to importance of each channel using two fully-connected (FC) layers and a sigmoid activation function.

5. The earthquake event classification method using an attention-based neural network according to claim 4 ,

wherein said performing of the excitation operation includes:

passing the compressed feature map through a dimension reduction layer having a predetermined reduction ratio; and

restoring the data passing through the dimension reduction layer to a channel dimension of the feature map.

6. The earthquake event classification method using an attention-based neural network according to claim 1 ,

wherein said performing of a nonlinear mapping by passing input data of the preprocessed earthquake data through the plurality of convolution layers having three or more layers further includes:

batch-normalizing the feature value of the preprocessed earthquake data in a first convolution layer and a last convolution layer among the plurality of convolution layers.

7. The earthquake event classification method using an attention-based neural network according to claim 6 ,

wherein said batch-normalizing of the feature value obtains an average and dispersion of each feature value and then converts regularized feature values while adding a scale factor and a shift factor thereto.

8. The earthquake event classification method using an attention-based neural network according to claim 1 ,

wherein said classifying of the earthquake event by regularizing the down-sampled feature value passes through first and second fully-connected (FC) layers, and the first fully-connected layer performs a drop-out regularization process in which each neuron is activated according to a probability value and applied to learning.

9. A non-transitory computer-readable recording medium, in which a computer program for executing the earthquake event classification method using an attention-based neural network according to claim 1 is recorded.

10. An earthquake event classification device using an attention-based neural network, comprising:

one or more processors configured to execute instructions stored in a memory, thereby configuring the one or more processors to:

preprocess input earthquake data by centering;

control a plurality of convolution layers having three or more layers to:

extract a feature map by nonlinearly converting the preprocessed earthquake data,

measure an importance value of a learned feature of the nonlinear-converted earthquake data based on an attention technique in which interdependence between channels of the feature map is modeled,

correct a feature value of the measured importance value through element-wise multiply with the learned feature map, and

perform down-sampling through max-pooling based on the corrected feature value; and

control first and second fully-connected (FC) layers to classify an earthquake event by regularizing the down-sampled feature value.

11. The earthquake event classification device using an attention-based neural network according to claim 10 ,

wherein each convolution layer is configured to:

nonlinearly convert the preprocessed earthquake data;

output a feature value of the nonlinear-converted earthquake data by measuring the importance value of the learned feature based on the attention technique in which interdependence between channels of the feature map is modeled; and

perform down-sampling through max-pooling based on the feature value.

12. The earthquake event classification device using an attention-based neural network according to claim 11 ,

wherein each convolution layer is further configured to:

compress the feature map of the nonlinear-converted earthquake data while maintaining global information at each channel;

adaptively recalibrate the importance value according to importance of each channel by using two fully-connected (FC) layers and a sigmoid activation function; and

correct the feature value of the measured importance value through element-wise multiply with the learned feature map.

13. The earthquake event classification device using an attention-based neural network according to claim 12 ,

wherein each convolution layer is further configured to:

pass the compressed feature map through a dimension reduction layer having a predetermined reduction ratio; and

restore the data passing through the dimension reduction layer to a channel dimension of the feature map.

14. The earthquake event classification device using an attention-based neural network according to claim 10 ,

wherein a first convolution layer and a last convolution layer are configured to obtain an average and dispersion of each feature value and then convert the regularized feature values while adding a scale factor and a shift factor thereto.

15. The earthquake event classification device using an attention-based neural network according to claim 10 ,

wherein a first fully-connected layer is configured to perform a drop-out regularization process in which each neuron is activated according to a probability value and applied to learning.