62 / 2023-03-31 14:16:22

A Neural Network Classification Method of Microseismic and Blast Events Based on Time-frequency Domain Analysis

microseismic monitoring,CNN algorithms,Classification,Blasting vibration signal

    The disasters caused by the concentration of ground pressure, such as roof fall, collapse and rock burst, do serious harm in mining safety and cause casualties and financial losses. Microseismic monitoring technology has been effectively used in monitoring the safety of mine, which can monitor the rock failure and burst of mine and give ealy warning based on statistical analysis of long-term microseismic events. Considering the impact of various signals generated from mine production operations and equipment systems, the identification of microseismic events is a key problem in application of microseismic monitoring system.

    The classification methods of microseimic and blast events of existing microseismic monitoring systems are mainly based on man. The Analysts, who need to be proficient in rock mechanics, seismology, and know mine conditions very well, have to select microseismic events from hundreds of events, with thousands of waveforms being identified one by one. While it’s difficult for mine enterprises to equip such talents, and it’s an overwhelming workload for analysts.

    As a binary classification problem, machine learning has been wildly used in classification of microseismic and blast signals. Feng Dai et al used I-CEEMDAN to decompose the signals and obtain singular values to train the classification model based on k-NN. Longjun Dong et al extracted features based on probability density functions and Fisher classifier was established. They also applied CNN in waveforms training, achieving great accuracy rate. Guoyin Zhang et al proposed an anti-noise classifier that combines CWT and CNN.

In this paper, we analyze the mine vibration signals collected by BSN microseismic monitoring system in time-frequency domain, where we can convert the time domain waveform into a two-dimensional time-frequency space, enabling energy analysis of the waveform simultaneously in time and frequency domains. More information and details of microseismic and blast signals can be obtained from the energy spectrum of time-frequency space, which helps to learn more differences between microseismic and blast signals, and achieve higher classification accurate rate.

    Then we carry out research on automatic classification technology based on Convolutional Neural Network (CNN), which are a class of feedforward neural networks that include convolution computation and have a depth structure. The three-dimensional time-frequency energy spectrum images of signals will be normalized, and will be used in classification model establishment. Over a thousand field data obtained in a metal mine are used for training and testing, and the accuracy of classification is larger than 85%, which can help obtaining more accurate microseismic events to analyze.