Damage identification using wave damage interaction coefficients predicted by deep neural networks

Abstract

The ever-increasing demand for efficiency and cost improvements in lightweight structures with guaranteed safety and reliability is leading to the application of a damage-tolerant design philosophy. Here, accurate knowledge of structural health is critical to avoid catastrophic failures. This knowledge can be obtained by using advanced structural health monitoring (SHM) systems. For thin-walled lightweight structures, methods utilizing guided waves generated by piezoelectric transducers are well suited. The interaction between the guided waves and potential damages can be described by so-called wave damage interaction coefficients (WDICs). These WDICs are unique for each damage and depend solely on its characteristics for a given structure. Therefore, the comparison of known WDICs with estimated ones allows drawing conclusions about the current structural state. In this paper, a novel damage identification method for plate-like structures based on a database of such WDICs is presented. Selected damages are simulated numerically with finite elements to generate WDIC patterns. However, these simulations are computationally highly demanding, thus only a very limited number of damage scenarios can be simulated. This study proposes an innovative technique to substantially enhance the resulting WDIC database by using deep neural networks (DNNs). These DNNs enable smart interpolations and allow not only predicting WDICs for previously unseen damages at low computational costs but also the discovery of knowledge about the complex relationship between damage features and WDIC patterns. A comparison to other machine learning algorithms clearly shows the superior performance of the utilized DNNs for interpolating complex WDIC patterns. The proposed damage identification method is verified using advanced time-domain simulations of a large aluminum plate. A statistical analysis of correct identification rates in a common three-sensor setting is employed for assessing the general performance. It is demonstrated that carefully identified DNNs enable to accurately replicate and interpolate complex WDIC patterns. Furthermore, it is shown that these predicted WDICs allow identifying damage characteristics with high confidence.

Keywords: Damage identification; Deep neural networks; Guided waves; Non-reflective boundaries; Structural health monitoring; Wave damage interaction coefficients.