Early Earthquake Detection Using Graph Neural Network

Earthquakes pose a threat to individuals,homes,and structural architecture.Seismic risk can be reduced with the help of early warning systems,which are set up to issue advance warnings of impending seismic activity.The success of these systems,however,depends on how quickly a warning is provided before the beginning of intense shaking.As a result,the use of deep learning models like GNNs,CNNs,and RNNs to detect earthquakes from seismic raw waveform data has emerged as a promising area of investigation.However,training times may increase due to deep learning models’ multilayered architecture and the necessity for many epochs.In addition,training the model with saturating nonlinearities can be challenging.To address these issues,batch normalization and layer normalization techniques are applied to each mini-batch and the input of the model to training epoch count should be decreased and achieve a stable distribution of activation values.This improves the model’s accuracy training and predictions.In this study,the authors implement three neural network-based prediction models trained using deep neural networks for earthquakes.For improved training speed and accuracy in seismic event prediction,we introduce a batch-normalized graph convolutional neural network(BNGCNN)that utilizes batch normalization before the activation layer.The model was evaluated on the California seismic dataset,and the results were compared to those of the graph convolutional neural network(GCNN)model.Our model outperformed the baseline on the same dataset by lowering the error rates for longitude/latitude,depth,and magnitude to 3 km,0.7 km,and 0.04,correspondingly.Our second model,known as BNGCNNAtt,is a batch normalization and attention mechanism-based graph convolutional neural network.The number of seismic stations and geographic regions are not constraints on the model’s ability to predict earthquake depth and magnitude.Data from three-component seismic waves recorded at many stations are fed into the CNN model,together with a magnitude range of ≥ 3.0 as input.The input waveform data undergoes preprocessing before the CNN’s convolutional layers are responsible for generating the feature map.The convolutional layers generate a feature map,and the attention mechanism focusses on the most relevant parts of that map.Training the model effectively and reliably requires an extra layer of added through batch normalization,which operates in batches.Finally,the GNN utilizes utilized the extracted features and event location data to produce precise event predictions.The BNGCNNAtt model was evaluated on two different datasets.We evaluated the performance of our proposed model through analysis of two independent datasets,one from Japan and the other from Alaska,both of which show different seismic dynamics.We compared our model’s efficiency to three baseline models,namely GCNN,BNGCNN,and GCNNAtt,using the root mean square error(RMSE)as the evaluation metric.The results showed that our proposed model produced remarkable RMSE values for magnitude prediction,measuring 2.8 and 4.0 for Alaska and Japan,respectively.Furthermore,depth prediction in Alaska and Japan,attained RMSE values of 2.87 and 2.66,respectively.These very low RMSE values indicate that our proposed model outperformed all three baseline models on both datasets,offering extremely accurate predictions for earthquakes of varying magnitudes.Our third proposed model is the Stacked Normalized RNN(SNRNN)model that uses ensemble learning and normalization technique to predict earthquakes effectively.SRNN uses a stacked-based ensemble method to ensemble Simple RNN,GRU,and LSTM models.The proposed model accurately predicts the magnitude,depth,and magnitude type of an earthquake event at any number of seismic stations in any location.SNRNN outperforms the three baseline models,Simple RNN,GRU,and LSTM,on a large-size Turkish dataset.The stacked ensembled RNN technique with layer normalization performs significantly better than batch normalization.The proposed model achieves 3.16,3.24,and 2.46 RMSE values to detect magnitude,type,and depth,respectively.The abstract describes a study that investigates the application of deep learning models in earthquake prediction.Three models are proposed and tested on various seismic datasets,with batch and layer normalizing approaches used to improve model training and prediction accuracy.A batch-normalized graph convolutional neural network(BNGCNN),a neural network that uses graph convolutions with both batch normalization and an attention mechanism(BNGCNNAtt.),and a stacked normalized RNN(SNRNN)model that uses ensemble learning and normalization techniques are among the models.The results show that all three models outperform baseline models,with the SNRNN model outperforming them all.