Explainable time–frequency convolutional neural network for microseismic waveform classification

Abstract

Geological hazards caused by rock failure severely threaten the safety of underground projects, and thus microseismic monitoring systems have been deployed to monitor the rock mass stability. However, due to implicit subseries patterns and sparse distinguishing features, automatic discrimination of the microseismic waveforms of rock fracturing remains a great challenge. Deep neural networks offer powerful learning ability, but the unexplainability of the neural network carries substantial risks to decision-making in safety warning. To this end, we propose an explainable convolutional neural network XTF-CNN that supplies both excellent classification performance and explainability. XTF-CNN consists of two major modules: 1) a dual-channel classification module that learns microseismic features from both the time and frequency domains and 2) an explanation module that demonstrates fine-grained and comprehensible results. Experiments are conducted using microseismic wave-forms collected from a deep tunnel project. The results indicate that XTF-CNN achieves superior classification performance over rival methods and significant comprehensibility.