CNN-ECA based classification of natural earthquakes and quarry blasting

The rapid and accurate identification of natural earthquakes and artificially blasted earthquakes is crucial for effective earthquake monitoring and early warning. We used waveform data from 5480 natural earthquake events and 4482 blasting events recorded by 110 seismic stations in a quarry in Utah, USA from January 2013 to August 2017, to construct a deep machine learning based CNN-ECA model and accurately and efficiently identify and verify these two types of earthquakes. Firstly, these data were preprocessed by removing mean, trend, instrument response removal, resampling (100 Hz), and bandpass filtering (1–20 Hz). Afterwards, the Fast Fourier Transform (FFT), Continuous Wavelet Transform (CWT), and Short Time Fourier Transform (STFT) methods were used to transform the time-domain data of 1453 natural earthquake events and 1103 quarry blasting events in 2013, obtaining four different types of training sample data: time-domain, frequency-domain (FFT results), and time–frequency domain (CWT and STFT results). Next, the four types of sample data were trained and tested using the Efficient Channel Attention Convolutional Network (CNN-ECA) and traditional Convolutional Neural Network (CNN). The results showed that the CNN-ECA model outperformed the CNN model in all four test samples. Especially when using time–frequency data converted through STFT and FFT as input, the recognition performance of the network model is more significant, with test set accuracies reaching 97.94% and 97.80%, respectively. Finally, the trained CNN-ECA model was used to validate and analyze the natural earthquakes and quarry blasting events recorded between 2014 and 2017. The results indicated that the combined use of FFT and STFT/CWT input data to jointly discriminate seismic events further improved the accuracy of earthquake type identification.