Accurate identification of microseismic waveforms based on an improved neural network model

Abstract

Microseismic waveform identification is the basis of rock burst risk assessment and rock burst warning in metal mines. This paper presents a method combining a genetic algorithm and neural network algorithm to classify the signals collected by a microseismic monitoring system during the mining process. After the analysis of the characteristics of the signals, six parameters, the duration time (DT), rise time (RT), ring number (RN), ring number of rise time (RRN), maximum amplitude (MA), and main frequency (MF), are selected for waveform recognition, that is, these six parameters are taken as the inputs of the waveform classification model. After establishing a fitness function, the numbers of the nodes in the two hidden layers were determined by a genetic algorithm. A 6–79-1 neural network waveform identification model was established, including one input layer, two hidden layers and one output layer. Only one blasting vibration waveform was misjudged as a rock mass fracture waveform when 114 samples were tested. Compared to the unordered multiple classification logistic regression method (UMCLRM), the improved neural network (INN) model had a higher classification accuracy. This model can effectively classify microseismic waveforms, mechanical vibration waveforms, electrical noise waveforms and blasting vibration waveforms. Additionally, the accuracy of waveform recognition is likely to be improved by using a two-layer hidden layer structure instead of by using a one-layer hidden layer structure. However, the complexity of the model increases. The waveform recognition model proposed in this paper is of great help to reduce the amount of manual work required to perform waveform recognition.