Evaluating Deep Learning Phase Picking Methods on EEWS Low-cost MEMS Accelerometers

In using EEWS based on low-cost MEMS accelerometers, calculating the P-phase is the most significant, considering that the P-wave is the first to occur in seismic signals during an earthquake. Providing a shorter time window in an extended warning time might also result in more false detections. The optimal period of 4 seconds for phase selection is crucial to the success rate of an EEWS. Conversely, a more significant time frame lowers false detections but offers a shortened warning time. Previous conventional methods have not been able to minimize false alarms. The STA/LTA approach achieved a test success rate of 100% but many non-earthquake signals. Thus, low-cost MEMS accelerometer sensors like P-Alert must be examined utilizing deep learning algorithms. The experiment indicated that the deep learning method using the stead pre-trained GPD (Generalized Phase Detection) outperformed the method previously implemented by P-Alert. Testing various seismic earthquake signals, GPD obtained an average timing error of 3.19 seconds for earthquakes and a 5-second time window. GPD with stead pre-trained contributes to a deep learning model with short-time errors and is practicable. These studies are believed to solve false alarms that occur during earthquakes.