Ensemble learning for inferring the time to failure of laboratory faults

In recent years, machine learning methods have made significant advances in laboratory earthquake prediction, opening up new directions for research into natural earthquake forecasting. However, most current studies primarily employ single shallow or deep learning models to fit the complex nonlinear relationship between acoustic emission data and fault instability states. This approach has certain limitations in fully capturing data complexity and enhancing model generalization. To address these challenges, this paper systematically compares the performance of Multi-Layer Perceptron (MLP) and Multi-Layer Perceptron with Bootstrap Aggregation (MLP-Bagging) models in the task of inferring the time to failure (TTF) of laboratory faults. The training and testing datasets are evaluated using multiple performance metrics, including mean absolute error (MAE), root mean square error (RMSE), mean squared error (MSE) and coefficient of determination (R2), enabling a comprehensive analysis of the fitting and generalization capabilities of both models. The results show that the MLP-Bagging model consistently outperforms the single MLP model across all statistical metrics, exhibiting lower prediction errors and higher R2 values on both the training and test sets. This finding is further validated in ensemble learning frameworks using LSTM + CNN as base learners. The Bagging ensemble strategy increases data diversity through bootstrap sampling and effectively reduces inference variance via model averaging, thereby significantly improving model stability and generalization. Although the inference accuracy of the models still requires improvement in regions near fault instability, ensemble learning methods provide a more robust and reliable approach for inferring fault instability states in laboratory settings. These results offer important reference value for related research areas, such as earthquake prediction.