A shallow Bayesian neural network with wavelet transform for impact localization in multi-material plate structures

Developing impact localization systems that are robust to material variability and computationally efficient remains a significant challenge for resource-constrained and self-powered structural health monitoring (SHM). The application of Bayesian regularization to shallow ANNs presents a promising avenue. However, its potential for creating an efficient model capable of generalizing across structures with extremely different material properties has not been previously fully explored in the literature. A shallow Bayesian-regularized artificial neural network (BRANN) in conjunction with wavelet analysis is proposed for impact localization in multi-material structures. The proposed model, which comprises two hidden layers, processes time-delay features of impact-induced guided waves. These features are extracted from piezoelectric sensor signals via wavelet transform. Experimental validation was conducted on plate structures with dissimilar material properties, specifically conventional aluminum and additively manufactured polylactic acid. The plate structures were integrated with an array of piezoelectric sensors to capture waveform signals under varying impact locations and energy levels. The performance of the BRANN was rigorously evaluated under three distinct sensor layouts. Results demonstrate the model's accuracy and robustness in achieving an average localization error below 4.4% across all tested impact energies, locations, sensor layouts, and material types. Furthermore, the model exhibits high computational efficiency, with an inference time of less than 8μs. Unlike complex deep learning models, the integration of a shallow Bayesian-regularized neural network with wavelet-transform offers a robust and computational efficient solution for resource-constrained applications.