Real-time prediction of drilling forces inside lunar regolith based on recurrent neural networks

Abstract

Currently, uncrewed lunar drilling and sampling using robots faces many technical difficulties. To ensure the efficient and reliable execution of lunar regolith sampling tasks, it is necessary to conduct research on robotic self-adaptive drilling strategies on the lunar surface. With the goal of developing autonomous uncrewed self-adaptive drilling strategies on the lunar surface, this paper presents a spatiotemporal feature fusion method for the real-time prediction of drilling forces inside a lunar regolith simulant. By extracting drilling state parameters in real time using a data acquisition system, a drilling load parameter predictor based on recurrent neural networks was trained to predict the penetration force and rotational torque during the next second accurately. Sensitivity analysis was performed on the data using random forests to quantify the importance of input variables. The proposed method provides a novel strategy for predicting drilling forces in real time and optimizes control parameters to guarantee safe and intelligent lunar drilling. Case study results demonstrate that (1) the coefficient of determination and mean absolute percentage error of the proposed spatiotemporal feature fusion method are 0.903 and 0.109, respectively, meaning it can provide accurate real-time prediction results. Additionally, (2) sensitivity analysis results demonstrate that the drilling force data from the previous 5 s have a significant impact on the drilling forces in the next 1 s. The drilling forces in the next 1 s are highly dependent on the state parameters in the previous 1 s. Additionally, the spatial relationships of the drilling state parameters are stronger than their time dependence. Practical application to the Chinese Chang’ E 5 mission for predicting the drilling forces of the robotic drill validates the effectiveness of the proposed method.