This paper proposes an efficient method for identifying artificial blasting waveforms based on the AlexNet convolutional neural network, which is designed earlier and still widely used today. This method directly uses event records as input data, which can simplify the data preprocessing, shorten the event determination time, and achieve a simple and fast effect. From the records of the Guangdong seismic network, this paper selects 312 artificial blasting events with M_L＞1.8 and 526 natural earthquake events with M_L＞1.4 that were manually analyzed and entered into the database. To achieve the best identification results, waveform within 30 km of the epicenter were selected, and after waveform preprocessing, such as trimming waveforms to a uniform length, waveform normalization, and removing abnormal information like calibration, square waves, sudden jumps, interference, and instrument faults, a total of 1840 valid waveforms were obtained. Among them, 1000 natural earthquake waveforms and 300 blasting waveforms were used to train the model, building an automatic blasting events classifier suitable for the Guangdong region. Additionally, this paper also studied the effect of changing the training set and validation set ratio during training and the training sizes on the accuracy. The results show that in this model, the accuracy reaches its optimum when the training set to validation set ratio is 8∶2, and when the number of training samples exceeds 600, the accuracy is higher than 95%. Finally, the well trained classifier was then tested by 540 waveforms from the Guangdong region, and it correctly identified 526 waveforms in less than 2 seconds, with an accuracy rate of 97.41%. Its precision, recall rate, and F₁-score for natural earthquake events are all greater than 0.98, while its precision, recall rate, and F₁-score of artificial blasting events are all greater than 0.90. All of these indicate that, on one hand, the size of the training sample needed for the model to achieve high accuracy is small, showing that this method is quite efficient. On the other hand, the AlexNet convolutional neural network model demonstrates higher adaptability to natural earthquake events with more training samples. With the input of more blasting events in the future, the model's recognition rate of artificial blastings will be further improved.

In conclusion, the AlexNet convolutional neural network can efficiently and accurately distinguish between natural earthquakes and artificial blasting in the Guangdong region. It can meet the requirement of timely and accurately removing artificial blasting events from the natural earthquake catalog to ensure the completeness and accuracy of the catalog, which is beneficial for regional strong earthquake prediction and seismic hazard assessment. Compared to manual work, this approach is more stable, accurate, and efficient. So the blasting classifier based on this model will save a significant amount of time and manpower for the work of Guangdong seismic network and provide support for the results of post-earthquake emergency response. Future research will focus on the practical application of this classifer, based on the data recorded by the Guangdong seismic network, continuously improving the accuracy and robustness of the classifer through constant testing, and applying it to practical earthquake early warning and daily seismic monitoring.