A data-driven hybrid recurrent neural network and model-based framework for accurate impact force estimation

Abstract

A novel hybrid deep learning technique is proposed, utilizing Recurrent Neural Networks (RNN) with four distinct sequential layers, designed to reconstruct unknown impact forces. The architecture integrates two interconnected sub-RNN structures, collectively referred to as IS-RNN, each consisting of the four-layer configuration. The first IS-RNN generates a transfer matrix, which serves as an input sequence for the second IS-RNN. To validate this approach, experiments were conducted using a rectangular carbon-fiber epoxy honeycomb composite panel, a structure commonly used in aerospace structures. A thorough analysis was performed, evaluating parameters such as signal length and solution methods. Comparative results between this technique and the Truncated Singular Value Decomposition (TSVD) method demonstrated strong alignment, with low percentage errors ranging from 1% to 2% when compared to actual impact forces. These findings highlight the proposed technique’s effectiveness and accuracy in impact force reconstruction.