A Normal Time-Frequency Transform Method for Automatic Recognition of Earthquake Event

  Objectives  In the era of earthquake big data, it is of great significance to develop efficient automatic and accurate seismic recognition algorithms. Convolutional neural networks (CNN) has played an effective role in seismic automatic recognition in seismology. However, in the environment of noise interfer‍ence, CNN identification accuracy is easily affected by noise and leads to decline. Generally, CNN meth‍ods eliminate the data with serious noise pollution through filtering and preprocessing or manual screening, so as to improve the accuracy of identification.

  Methods  In view of the problem of seismic event recognition, we propose a new method, normal time-frequency transform (NTFT) and CNN are combined for realizing automatic and accurate identification of seismic signals. The NTFT+CNN model has good robustness and does not require manual pre-processing and denoising of seismic data in advance. Tak‍ing 19 884 seismic events and 17 640 noise data randomly sampled from the Yingjiang County of Yunnan prov‍ince, China, the comparison between CNN and NTFT+CNN was conducted from the respect of large samples and small samples.

  Results  In the experiments of the three stations, the recognition accuracy of CNN model is 93.10%, and the recognition accuracy of the NTFT+CNN model is 97.80%, which means that the introduction of NTFT reduces the recognition error rate by three times. It shows that NTFT+CNN model can recognize the seismic signal more effectively under the condition of noise interfer‍ence. At the same time, the epoch in CNN model is 29 and that of in NTFT+CNN is 18, which shows that the convergence in NTFT+CNN model is quicker and more stable. In the experiment of a single station, the comparison of three typical noise conditions was investigated. The recognition accuracy of CNN model was 95.98%, 91.56% and 90.36% in weak background noise, strong background noise and similar waveform noise, respectively, while the recognition accuracy of NTFT+CNN model was 99.80%, 97.21% and 98.50%, respectively, which further verified the recognition effect of NTFT on noise and seismic signal, the correctness of model results. The NTFT+CNN model was applied to identify the seismic-impulse noise data shared by the Southern California Earthquake Data Center. Comparing to CNN, NTFT+CNN model has more advantages on convergence speed, recognition accuracy and the number of training samples required.

  Conclusions  NTFT provides a new and efficient way for the recognition of seismic signal in neural network structure.