Training Structures as Distributed Sensors: A case study of impact force identification using neural operators

Structural sensing data not only provides insights into a structure’s behaviors but also contains valuable information about its surrounding environments. The concept of “Structures as Sensors” has introduced the idea of enabling structures to be environment-aware and serve as a sensing platform. In this paper, we propose extending this concept to “structures as distributed sensors” (S2DS), utilizing learning-based methods to infer surrounding environments from structural vibration data. Our approach addresses the challenge of perceiving environmental events not present in the training dataset, particularly highlighting high-resolution solutions – providing solutions across densely distributed locations, with limited deployed sensors. As a proof of concept for S2DS, we present a study on impact force identification using Fourier Neural Operator (FNO) with a proposed spatial upsampling mechanism. FNO is designed to infer spatiotemporal source functions from structural responses, allowing for the simultaneous localization and reconstruction of impact forces within a unified framework. We also use a continuous distribution representation of single-point impact forces, which effectively tackles the issue of training on highly sparse data while ensuring precise localization. The validation of our framework on a cantilever beam and a wind turbine blade demonstrates accurate identification of both the location and time history of impact forces, irrespective of the localization resolution. Importantly, our framework delivers satisfactory outcomes with a limited training dataset, highlighting its potential for structural health monitoring and environmental perception of engineering structures in various scenarios with limited sensors and training resources.