Intelligent Perception for Impact Phase in Hard-Target Penetration based on Time-Series Retained Neural Networks

Abstract

During high-speed penetration, missile-borne impact sensing systems face complex conditions involving multiple materials, varying thicknesses, and variable velocities. The signals acquired by accelerometers exhibit significant differences in morphology, making real-time and accurate identification of the exit point difficult. This paper addresses the engineering constraints of missile-borne intelligent microsystems by proposing an intelligent perception method for impact phase in hard-target penetration based on time-series retained neural networks. At the sensor data preprocessing level, a refined filtering and dual normalization mechanism targeting highly correlated intervals is designed to eliminate differences in amplitude dimensions across operating conditions and smooth the error surface. At the network architecture level, a strategy of removing the global pooling layer and introducing a temporal flattening operation is proposed, hard-coding the temporal location information of the impact event into the feature arrangement structure, achieving end-to-end accurate regression from the original one-dimensional waveform to the exit point coordinates. Using metal steel plates and thick concrete walls as penetration targets, the system was validated under 18 different working conditions with three impact velocities and three target thicknesses. Results show that the proposed method achieves a target thickness classification accuracy exceeding 95%, with a maximum mean absolute error of 4.56 mm for steel plates and 19.5 mm for concrete. With only about 5.2 × 10⁴ model parameters, it can fully cope with the resource constraints of embedded processors, effectively enhancing the accuracy and real-time performance of the intrusion sensing system in complex dynamic environments while maintaining a lightweight design.