Slope failure level prediction using a hybrid convolutional long short-term memory network based on microseismic monitoring data

The gradual deterioration of slope rock masses could lead the slope prone to fail under mining-induced disturbances, which poses significant threats to personnel and equipment safety. Accurate prediction of slope failure severity levels constitutes a critical foundation for effective disaster prevention and mitigation. This work implemented the microseismic monitoring technology to establish a slope failure information database, construct an early warning model, and conduct predictive research on slope failure severity levels. This study presents a three-level classification method for slope failure. The number of microseismic events, microseismic energy, seismic moment, cumulative apparent volume, rainfall, and aggregation index were selected as early warning indicators via correlation analysis. These early warning indicators build an early warning index system for slope failure severity levels. This study introduces an attention mechanism and an improved gating mechanism to establish a hybrid convolutional long short-term memory (LSTM) model for slope early warning. The model enhances feature channels relevant to the current task, increasing the importance of early warning indicator features by 50%. By integrating the input gate, forget gate, and output gate structures into an update gate and reset gate, the internal architecture of the model is streamlined, improving its computational efficiency by 15%. The proposed hybrid model achieves an accuracy of 80.9% in predicting slope failure severity levels, representing improvements of 5.9% and 11.7% over conventional convolutional and LSTM models, respectively. This study provides an efficient and accurate technique for predicting slope failure severity levels, offering significant implications for slope safety research in engineering applications.