Application of Deep Learning in Microseismic Waveform Classification: A Case Study of The Yebatan Hydropower Station Project

Deep learning plays an important role in automatic classification of microseismic (MS) waveforms, but there is still a problem of low recognition accuracy based on small training samples in engineering applications. We propose an enhanced convolutional natural network (ECNN) based on ACGAN structure for MS waveforms classification, where the generator is used to synthesize samples of specified type and the discriminator is used to identify class and authenticity. The generator and the discriminator are confronted with each other through alternate training, so as to ensure the accurate classification of MS waveforms with small training samples. Then, relying on the Yebatan hydropower project, 15503 microseismic waveforms are processed to study the impact of training sample changes on ECNN and traditional CNN model. The results show that the classification accuracy of the two models tends to be stable with more than 1024 samples, and decreases rapidly with less than 512 samples; ECNN method has higher classification accuracy than traditional CNN method under the same training samples, and the accuracy increases more significantly with the reduction in sample size. The research results can reduce the difficulty of modeling waveform classification and help the application of deep learning in MS monitoring.