Traditional source localization techniques like matched field processing (MFP) rely on physics-based models of known acoustic environments, which can be impractical for complex systems in structural acoustics. In contrast, machine learning (ML) methods are “model-free,” leveraging patterns in data to predict source locations. This presentation compares MFP and neural network-based ML approaches for localizing impact sources on a 6.4-mm thick circular aluminum plate with diameter 914 mm. The plate was excited by a 12.7-mm-diameter stainless-steel ball bearing dropped from 76 mm above the surface. A linear array of 14 microphones, spaced 51 mm apart and located 92 mm above the plate, recorded the acoustic response of the impact. Because experimental data can be time- and resource-intensive to collect, finite element simulations generated acoustic responses for 8640 source locations in the 5–20 kHz bandwidth. This simulated data was used to train and validate two neural networks: a feedforward neural network using cross-correlation lags and a convolutional neural network using spectrograms. Both models were tested, achieving localization errors down to 0.3 cm (simulated) and 1.0 cm (experimental). ML approaches are also evaluated in terms of data efficiency, computational speed, and robustness to noise and signal variability. Sponsored by a SMART Scholarship.
King et al. (Wed,) studied this question.