GAN-LSTM Joint Network Applied to Seismic Array Noise Signal Recognition

• Published in: Applied sciences Vol. 11; no. 21; p. 9987
• Main Authors: Li, Jian, Hei, Dongwei, Cui, Gaofeng, He, Mengmin, Wang, Juan, Liu, Zhehan, Shang, Jie, Wang, Xiaoming, Wang, Weidong
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.11.2021
• Subjects: Arrays; Beamforming; Classification; Data processing; Datasets; Deep learning; Earthquakes; Emergency communications systems; GAN; Long short-term memory; LSTM; Machine learning; Monitoring; Neural networks; Noise reduction; noise signal recognition; Nuclear explosions; Parameter estimation; Phases; Seismic activity; seismic array; Seismic exploration; Signal processing; Signal quality; Signal to noise ratio; Waveforms; United States > US; Oklahoma
• ISSN: 2076-3417; 2076-3417
• DOI: 10.3390/app11219987

The purpose of seismic data processing in nuclear explosion monitoring is to accurately and reliably detect seismic or explosion events from complex ambient noises. Accurate detection and identification of seismic phases are of great significance to the detection and parameter estimation of seismic events. In seismic phase identification, discriminating between noise signals and real seismic signals is essential. Accurate identification of noise signals helps reduce false detections, improves the accuracy of automatic bulletins, and relieves the workload of analysts. At the same time, in seismic exploration, the prime objective in data processing is also to enhance the signal and suppress the noises. In this study, we combined a generative adversarial network (GAN) with a long short-term memory network (LSTM) to discriminate between noise and phases in seismic waveforms recorded by the International Monitoring System (IMS) array MKAR. First, using the beamforming data of the array as the input, we obtained the signal features of seismic phases through the learning of the GAN discriminator network. Then, we input these features and trained the joint network on mixed seismic phase and noise data, and successfully classified seismic phases and noise signals with a recall of 95.28% and 97.64%, respectively. Based on this model, we established a real-time data processing method, then validated the effectiveness of this method with real 2019 data of MKAR. We also verified whether improved noise signal identification improves the quality of phase association and event detection.