Earthquakes are among the most devastating natural disasters, resulting in significant loss of life and extensive damage to property. These events occur primarily due to the collision of tectonic plates, which generates stress and energy that is ultimately released as an earthquake. Identifying critical zones that are particularly susceptible to seismic activity is essential, as it allows for the implementation of effective
mitigation measures.
To this end, Earthquake Early Warning (EEW) systems are crucial tools for minimizing the impact of earthquakes. Research has demonstrated their effectiveness in reducing hazards. EEW encompasses various components, including source location warnings, intensity warnings, and magnitude warnings.
“Research on EEW has shown that if the earthquake warning time is 3s, the casualty rate can be reduced by 14%; if the time is 10 s, the casualty rate can be reduced by 39%; and if the time is 60 s, the casualty rate can be reduced by 95%.“
Numerous researchers have proposed various methods for Earthquake Early Warning (EEW), including empirical, wavelet analysis, and magnitude warning methods. While these approaches have demonstrated effectiveness to a certain degree, they often experience a significant level of dispersion.
With advancements in computational techniques and the accumulation of extensive earthquake data, there has been a notable shift toward the application of deep learning (DL) techniques in the field of EEW. The convolutional neural network (CNN) is among the most widely utilized DL techniques. Several researchers have employed it, and the results indicate an impressive reduction in error rates by up to 80%.
A recent study conducted by a research team from the Software College of Northeastern University in Shenyang, China, employed a CNN model known as EEWMagNet for EEW purposes. The study utilized three-component waveform data with a sampling frequency of 100 Hz, recorded by the China Earthquake Networks Center (CENC) from 1,104 stations between 2009 and 2017. For the modelling process, the dataset was divided into 80% for training and 20% for testing. The study also aimed to identify the most cost-effective input length by varying the input duration from 0 to 10 seconds.
The analysis indicates that optimal model performance is achieved when the input length reaches 7s, with recorded accuracy levels reaching 90%. This enhancement can be attributed to the increased data length, which enables the model to extract a greater number of features essential for accurate magnitude estimation. However, beyond the 7s threshold, performance begins to decline. This reduction is likely due to the introduction of excessive data that fails to provide additional key features and may introduce noise. Consequently, the decision was made to establish 7 seconds as the most cost-effective input length for the EEWMagNet model.
Given the limited informativeness of the available data for Earthquake Early Warning (EEW), the authors opted to implement the DenseBlock architecture for feature extraction. This choice allows for rigorous feature extraction and fusion, thereby enhancing the warning system’s accuracy. While the application of DenseBlock increases network complexity, the Bottleneck module has been incorporated to mitigate computational costs effectively.
The bottleneck configuration in EEWMagNet includes a 1×1 convolutional layer for dimensionality reduction, a Batch Normalization layer to enhance convergence, and a ReLU activation function. To mitigate overfitting and reduce computational costs, a transitional module is integrated between dense blocks. While dense blocks are effective for feature extraction, they struggle to capture temporal correlations, which are crucial for improving the network’s performance.
Recurrent Neural Networks (RNNs) are typically used for processing temporal data, but they face limitations when handling longer sequences due to issues with short-term memory. To address this, a Multi-Head Attention structure is implemented, allowing for parallelized computation and effectively considering the correlation of features.
The EEWMagNet model is designed and implemented with the following components: a Dense block layer featuring a bottleneck, a transition layer, a Multi-Head Attention layer, and an epicenter distance fusion layer. It consists of four Dense block layers, with 6, 12, 24, and 16 convolutional layers, respectively, each using 32 convolutional kernels. The output from these layers generates 256, 512, 1024, and 1024 feature channels. The epicenter distance is integrated with the average pooling output for magnitude classification. Finally, the model employs two fully connected layers to classify the output feature vectors.
The study evaluates two models: (i) the full EEWMagNet network and (ii) the EEWMagNet network excluding the Multi-Head Attention structure. Performance was assessed using the accuracy metric. The results indicate a modest improvement in accuracy of 2.15% for the EEWMagNet model compared to the version without the Multi-Head Attention structure. This addition significantly enhances the model’s effectiveness for Earthquake Early Warning (EEW) classification tasks by improving its ability to capture sequence correlations within seismic data, thereby making the EEWMagNet network more adept at processing such information.
The study introduces the EEWMagNet model for classifying earthquake magnitudes and providing early warnings. EEWMagNet employs convolutional neural networks to extract features from single-station three-component waveform data and includes epicentral distance to improve its performance. A key contribution is the development of this deep learning network, which offers faster and more accurate magnitude classification compared to traditional methods, even with shorter waveform data.
(Source: https://doi.org/10.1016/j.aiig.2023.03.001)
Reference
Meng, F., Ren, T., Liu, Z., & Zhong, Z. (2023). Toward earthquake early warning: A convolutional neural network for rapid earthquake magnitude estimation. Artificial Intelligence in Geosciences, 4, 39-46.
Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on image data augmentation for deep learning. Journal of big data, 6(1), 1-48.
Ioffe, S. (2015). Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167.
Vaswani, A. (2017). Attention is all you need. Advances in Neural Information Processing Systems.