With the improvement of earthquake monitoring capabilities and the surge of monitoring data, the research on seismology has entered the era of big data. Especially with the increase of mining blasting, engineering demolition, military construction and other activities, seismic stations will collect a large number of natural and artificial blasting waveform records. Accurately and quickly identifying artificial blasting and natural earthquakes from waveform events has become one of the focuses of earthquake warning and prediction research. Numerous scholars have conducted in-depth researches on earthquake event classification and recognition. The use of convolutional neural network （CNN） technology for earthquake event detection and classification is currently one of the research hot-spots, but one of the key challenges is how to capture the different features of artificial blasting and natural earthquakes.

In order to further study the application of CNN in the field of earthquake event automatic detection and improve the efficiency of event automatic detection, a study was conducted on the classification and recognition of natural earthquakes and blasting events based on CNN, with 12 936 artificial blasting micro-seismic records and 13 215 natural micro-seismic records with magnitude M_L1.3−3.0 as the research objects.

Firstly, the seismic waveforms are preprocessed. The original seismic waveforms are filtered using a band-pass filter with a range of 1−30 Hz to remove long-period interference components, resulting in distinct P- and S-wave records. Based on this, P-wave identification is performed using short-term/long-term average （STA/LTA） algorithm, with STA duration set as 0.2 seconds, LTA duration set to 1 second, and threshold size set to 2. The waveform from 20 seconds before the first arrival time to 100 seconds after the last arrival time was taken as the screening result for this record, resulting in 12 132 effective natural earthquake screening records and 11 721 artificial blasting screening records.

Secondly, the S-transform is applied to obtain the S-transform spectrum of the preprocessed seismic signals, and the S-spectrum energy curve that varies with frequency is calculated. The S-spectrum energy curve can clearly depict the frequency and energy changes of seismic signals, and better characterize the characteristics of the original signals.

Then, based on the classic LeNet5 model, a CNN network model was constructed, which includes one input layer, three convolutional groups consisting of three convolutional layers and three pooling layers, one fully connected layer, and one output layer. In order to reduce resource loss and time consumption, and improve operational efficiency, the three channel RGB image of the S-spectrum energy curve is converted into a grayscale feature map with a size of 32×32 pixels, which is used as input for CNN. The CNN model is trained using the training set to obtain the optimal CNN model parameters.

Finally, algorithm testing is conducted based on the trained CNN model. Randomly extract a certain proportion （50%−90%） of the preprocessed seismic record dataset as training data, and use the remaining data for testing to verify the recognition accuracy of natural earthquakes and artificial blasting events. The tests show that the more training samples there are, the better the classification and recognition performance. When the training sample ratio is 90%, the average recognition accuracy is 97.57%.

The algorithm performance was tested using the ten-fold cross validation method, which was repeated 100 times. The average recognition result was taken, and the recognition accuracy reached 97.80%. The sensitivity （SE） and specificity （SP） values of the classification performance indicators were close, which indicates good recognition performance.

To further test the effectiveness of the S-spectrum energy curve as a feature for seismic signal classification and recognition, the S-spectrum, wavelet spectrum, short-time Fourier transform （STFT） spectrum, and fast Fourier transform （FFT） spectrum were used as inputs for CNN model training and testing. The results showed that compared with other signal spectra, the S-spectrum energy curve can intuitively reflect the energy magnitude and changes of each frequency component of the signal, with a higher recognition accuracy of over 97%.

The experimental results show that the S-spectrum energy curve can serve as an effective and reliable basis for classifying natural earthquakes and artificial blasting events, and the CNN model in this paper is reliable, with good stability and accuracy.

As it should be, using simple binary classification problems such as natural earthquakes and artificial blasting alone is not enough to describe the complexity of earthquake event classification and recognition. In the next step of research, we will collect sample data from different regions and earthquake events, optimize the structural model of CNN training networks, and make the model for earthquake event detection and recognition more accurate and the recognition effect more intelligent.