Abstract:

To address the problems of low recognition accuracy, poor generalization, and high deployment cost in traditional algorithms for mine microseismic waveform classification, a multi-scale residual convolutional neural network (MSRCNN) was designed for the complex time-frequency structures and significant scale variations of microseismic waveforms. Multi-scale convolutions were used to capture local details and global trends of microseismic waveforms. Residual connections were introduced to enhance deep feature transmission and gradient stability. The MSRCNN was conmbined with a bidirectional long short-term memory network (BiLSTM) and a self-attention mechanism to construct the MSRCNN-BiLSTM-Attention model. A weighted loss function combining cross-entropy loss and supervised contrastive learning was used for training optimization. The model was validated using data from multiple mine microseismic monitoring projects. The results show that the constructed model performs significantly better than traditional image recognition models in terms of accuracy, recall, and precision. High accuracy is maintained in transfer tests across different mines. Good robustness and generalization ability are also shown. The MSRCNN module is the core source of model performance. When used alone, it achieves recognition accuracy close to that of the complete model, and it also has fewer parameters. Therefore, it is suitable for deployment in real-time mine monitoring systems with limited computing resources. The complete model has more parameters, but it achieves the best recognition performance. It is more suitable for high-risk scenarios, such as deep high-stress areas, rockburst-prone areas, and areas with frequent microseismic triggering. More reliable recognition ability is provides for microseismic early warning.

Key words: multi-scale residual convolutional neural network (MSRCNN), mine microseismic, image recognition, bidirectional long short-term memory network (BiLSTM), self-attention mechanism