Leveraging nonlinear hysteretic response for structural characterization

Nonlinear resonance ultrasonic spectroscopy (NRUS) is a nondestructive testing method that has been demonstrated to be highly sensitive to the presence of damage, especially in cases of progressive microcracking such as in mechanical or thermal fatigue. In controlled laboratory experiments, a characteristic strain-dependent hysteretic NRUS response can be determined via tracking of resonance frequencies. However, such results can be limited in their scope and reliant on controlled excitation sources. In recent NRUS experiments, we have incorporated machine learning to help address these limitations. Here, we present the nonlinear response of samples under pseudo-random excitation and discuss neural network predictions of structural status and experimental conditions, through training on 3-D frequency response functions. Using sparse sampling of a dataset which included mode- and stress-dependent nonlinear responses of a damaged component, structural status predictions using convolutional and graph neural network architectures had an accuracy of 85–90% in many cases, nearly four times better than random chance. This work is motivated by ongoing development of nuclear microreactors, which operate under more extreme conditions than conventional reactor designs, necessitating new control systems and sensor technologies to monitor for structural anomalies on the component level.