Spatial deduction of mining-induced stress redistribution using an optimized non-negative matrix factorization model

Abstract

Investigation of mining-induced stress is essential for the safety of coal production. Although the field monitoring and numerical simulation play a significant role in obtaining the structural mechanical behaviors, the range of monitoring is not sufficient due to the limits of monitoring points and the associated numerical result is not accurate. In this study, we aim to present a spatial deduction model to characterize the mining-induced stress distribution using machine learning algorithm on limited monitoring data. First, the framework of the spatial deduction model is developed on the basis of non-negative matrix factorization (NMF) algorithm and optimized by mechanical mechanism. In this framework, the spatial correlation of stress response is captured from numerical results, and the learned correlation is employed in NMF as a mechanical constrain to augment the limited monitoring data and obtain the overall mechanical performances. Then, the developed model is applied to a coal mine in Shandong, China. Experimental results show the stress distribution in one plane is derived by several monitoring points, where mining induced stress release is observed in goaf and stress concentration in coal pillar, and the intersection point between goaf and coal seam is a sensitive area. The indicators used to evaluate the property of the presented model indicate that 83% mechanical performances have been captured and the deduction accuracy is about 92.9%. Therefore, it is likely that the presented deduction model is reliable.

Keywords

• Machine learning
• Underground construction
• Monitoring
• Mining-induced stress
• Prediction