Novel Transfer Learning Framework for Microseismic Event Recognition Between Multiple Monitoring Projects

Deep neural networks have achieved great success in microseismic event recognition tasks in rock engineering. However, a well-trained model on one project may not deliver satisfactory results on other new projects due to the different geological conditions, construction methods and monitoring configurations. In this study, the differences of microseismic monitoring in four monitoring project cases, including the geology conditions and monitoring settings, are compared to better understand the challenges of signal recognition by deep learning, and then the drawbacks and unavailability of current algorithms for microseismic monitoring scenarios are discussed. To address the challenges, a unified transfer learning strategy is proposed to perform knowledge transfer for microseismic recognition between multiple projects with varying difficulties, and it consists of a novel unsupervised two-level adversarial adaptation method and a semisupervised active learning mechanism; thus, it operates in a semisupervised fashion. The results experimentally demonstrate that a well-trained model cannot generalize well to other rock projects without performing transfer learning. This framework is evaluated on four real-world collections of monitoring recordings and compared with other benchmark methods. The results show that it outperforms all counterparts and is able to conduct knowledge transfer with varying difficulties. Moreover, the proposed active learning mechanism significantly reduces the manual labeling costs for both hard and easy tasks. The simulations suggest that we have preliminarily achieved deep model reuse in signal recognition of microseismic monitoring.