Prediction of Missing Frequency Response Functions Through Deep Image Prior

Vibration analysis is crucial when designing and monitoring resonant structures. The characterization of vibrational properties in mechanical systems, e.g. machinery or musical instruments, can indeed detect noise sources and damages. Several methods can retrieve these parameters starting from a set of measurements. The level of detail in the estimate mostly depends on the amount and distribution of points acquired over space. A potential issue for these techniques consists in the presence of regions over the object where sensors cannot be attached. In this case, an interpolation scheme with a suitable prior of the data model should be devised. We propose here to predict the missing vibrational data within the framework of image inpainting and apply a fully data-driven method based on Deep Image Prior, which allows to capture the prior inside data without the need of a dataset. The performance is assessed in the case of violin top plates. The proposed method proved to better predict data, in particular resonances, for points close to the boundary, whereas a baseline based on Thin Plate Splines fails, due to the reduced number of available samples.