A novel microseismic classification model based on bimodal neurons in an artificial neural network

Abstract

Microseismic monitoring systems deployed in deep underground engineering can capture massive waveform signals in real time. However, some noise signals are highly deceptive and similar to microfracture signals. This problem usually requires engineers to compare the signal characteristics from different domains and poses a challenge for the rapid detection of microseismic data. To address this problem, we establish a new waveform multiclassification model based on the multimodal feature extraction technology. The proposed model constructs a dual-branch convolutional neural network to learn the bimodal features from the time and frequency domains. Furthermore, the attention mechanism (channel and spatial attention) with skip connection is used to increase the focus on the prominent features of the waveform and reduce the interference caused by irrelevant information. We conduct experimental studies using microseismic data from deep tunnel projects under complex geological conditions and classify the recorded data into three target types: microseismic, blast, and noise waveforms. The results show that the proposed approach achieves good multiclassification performance for noise signals highly similar to microseismic signals as well as for noisy microseismic signals with different signal-to-noise ratios. In summary, this study lays the foundation for further exploration of the potential application of multimodal fusion in seismic engineering and various rock engineering fields.