A Novel Transfer Learning-Enhanced BiLSTM-DCNN Architecture for Mine Microseismic Signal Identification with Small Data Set

The underground mining environment produces highly heterogeneous microseismic (MS) signals. Accurately identifying rock failure signals is crucial for source localization and failure mechanism analysis. However, data scarcity during initial monitoring stages severely limits recognition accuracy. Existing methods perform poorly under small-sample conditions, with low identification rates and weak generalization. To address this, this study proposes a Transfer-learning-enhanced Bidirectional LSTM and Deep Convolutional Network (Tr-BiLSTM-DCNN) model. It uses Mel-spectrograms for signal characterization, combining BiLSTM's bidirectional temporal modeling with DCNN's multi-scale spatial feature extraction to build a spatiotemporal feature representation. Training involves a two-phase strategy: pretraining and hyperparameter optimization on large cross-mine datasets, followed by domain-adaptive fine-tuning via transfer learning on the target mine's small-sample data. The model achieves 94.44% test accuracy under small-sample conditions, an 80.85% improvement over non-transfer baselines, and outperforms conventional CNN and LSTM methods. It provides an intelligent few-shot learning solution for mine MS monitoring, showing strong potential for engineering applications in dynamic disaster early warning.