Recognition of seismic pulses using CNN models for zero-based segmented mini-signals

Structures require higher seismic demands for typical seismic pulses, particularly strong-velocity pulse motions that can adversely affect long-period structures. As proper pulse classification is essential in earthquake engineering, this study proposes a deep-learning-based method that uses standardized mini-signal images as inputs, resulting in model recognition accuracy of 99.86 %. A pulse counting standard is developed to effectively extract pulse features based on the pulse content output from the training model. The pulses are divided into two categories—primary and secondary—to be counted separately, and the number of pulses and the period are determined. The original signal of the 171th seismic event in the PEER NGA-West2 database was counted, and the record was found to contain two main velocity pulses with periods of 2.64 and 2.60 s. Comparing the proposed method with the two benchmark methods, the results show that the method can extract the pulse positions without limiting itself to the predefined mathematical model, adapt to the shape of the original signal, and extract seismic pulses more accurately. Furthermore, this method is suitable for the data extraction of multiple pulses. • A method for identifying pulse-like ground motions using deep learning is presented. • Pulse counting standards based on the inverse normalization method were proposed. • The method can extract pulses, even multiple pulses, from the original signal.