Speaker
Description
Structural intensity provides valuable insights into vibrational energy flow and dominant transmission paths within mechanical structures. For thin-walled plates, it can be derived from the transverse displacement field and its spatial derivatives. However, obtaining reliable estimates from experimental data is difficult, since numerical differentiation amplifies measurement noise, especially at higher derivative orders. This contribution introduces a machine learning framework based on a physics-informed deep operator network that predicts structural intensity directly from noisy vibration measurements. Rather than differentiating the raw measurements by finite-difference schemes, the approach constructs a smooth, fully differentiable neural network surrogate representation of the displacement field, from which higher-order spatial gradients are obtained through automatic differentiation. This strategy avoids the noise-induced instabilities inherent in numerical differentiation and yields robust predictions of structural intensity from experimental data. The methodology is evaluated on two test cases: an analytical benchmark problem involving a simply supported plate, and laser Doppler vibrometry measurements of a plate structure. The results demonstrate accurate reconstruction of displacement and structural intensity fields over a broad frequency range, capturing both energy-flow magnitudes and directional patterns with high fidelity. Compared with standard numerical differentiation techniques, the proposed data-driven approach consistently provides higher robustness to noise and increased predictive accuracy, supporting its applicability in experimental structural dynamics and noise control engineering.