Making a brain on a structure: a conceptual study of elastic wave field neural networks for structural health monitoring

Abstract

The purpose of this study is to develop a novel concept of smart structural systems that can recognize their own structural integrity by an embodied high density sensor network. Over the past two decades, sensor networks for automatic inspection application have been intensively investigated, and it has now become reasonable to deploy over 1000 sensor nodes in a single structural system. It would be certain, however, that the current approaches that require rich electronics and wireless communication at each sensor node will reach its limit due to huge amount of data overwhelming the network capacity and centralized computing resources. In this study, we propose a new approach to make a breakthrough in both communication and computation for such high density sensor networks of the next generation. In our approach, a number of sensor nodes with simple functions are embedded in the structure, each of which reacts to the elastic waves propagating through the structure by applying a force to the structure after a simple nonlinear transformation. This allows the whole nodes to be mutually coupled through the medium of elastic waves, forming a neural network that incorporates the dynamic characteristics of the structure as the coupling weights. In this paper, we present a possible realization of our concept with basic formulations, and present numerical simulations to examine how the proposed network behaves under a single frequency input. It is presented that the network exhibits a bifurcation in its asymptotic behavior from modulated response to steady-state depending on the structural conditions.

Physical Reservoir-Based Health Monitoring of a Structure with Nonlinear Attachments

“…Among the parities from the 1st to the 5th, it is found that the 2nd parity is the most sensitive to damage. This is the same as the result reported in the previous study (Masuda et al, 2023). The reason why the 1st parity is less sensitive than the 2nd parity may be due to the simplicity of the task, which could introduce the trained PRC some undesirable tolerance to the parameter variation.…”

“…Figure 1 illustrates the concept and entire workflow of the PRbased SHM proposed by the authors (Masuda et al, 2023).…”

“…The learning task used in the presented toy problem is the same as the previous work (Masuda et al, 2023), which was quoted from the reference (Coulombe et al, 2017). The learning task is defined as the nth-order parity function.…”

“…The idea is to utilize the dynamics of the target structure as a physical reservoir (PR) to perform some or most of the data processing and computation in SHM by using the structure itself as a computational resource. As a preliminary study, a simple nonlinear oscillator network was used as a model of the target structure in the previous paper (Masuda, Takashima, & Sakai, 2023), and a PRC system was defined by attaching input and output layers to the structural model as a PR, then a toy problem of learning a nonlinear function was conducted. It was demonstrated that changes in mechanical properties due to structural damage can be detected as an increase in the PRC output error.…”

“…In this study, we conduct similar investigation on a different target structure that is modified from the one used in the previous study (Masuda et al, 2023) to be slightly more realistic. There are two points of modification: first, it was assumed in the previous paper that the target structure consists of multiple oscillators with strong nonlinearity connected in a line, whereas the actual object structure is to be linear.…”