Multiple Dynamic Impact Signal Identification Method Based on Lightweight Neural Network With Acceleration Sensor

Multiple dynamic impact signals are widely used in many engineering scenarios. In the process of penetrating multilayer hard targets, the overload signals from multiple dynamic impacts are difficult to identify due to “signal adhesion” problems caused by nonlinear strong disturbances. For this problem, a customized lightweight neural network framework is proposed here. First, the characteristics of multiple impact signals and their disturbances are analyzed by dynamics modeling and continuous wavelet transform. Second, considering these characteristics with SqueezeNet, a simplification rule for a network redundancy structure is proposed. A customized residual feedback structure and attention mechanism are nested within the structure, thus improving feature recognition performance with less computational complexity. Third, experimental results show that the proposed method maintains comparable excellent feature recognition results to the complex GoogleNet under all test conditions with varying impact speeds, while its metrics for computational complexity are even lower than those of SqueezeNet and other classical lightweight networks by 23%–61.6%. Therefore, the proposed method is of great practical value for penetrating ammunition and other weak hardware application platforms that require accurate identification of multiple dynamic impact signals.