Spiking Neural Networks for Energy-Efficient Acoustic Emission-Based Monitoring

Acoustic emission (AE) is one of the most effective nondestructive testing (NDT) techniques for the identification and characterization of stress waves originated at the uprising of acoustic-related defects (e.g., cracks). To this end, the estimation of the time of arrival (ToA) is crucial. In this work, a novel processing flow which shifts the identification process from the time to the time-frequency domain via wavelet transform (WT) is proposed, allowing to better capture transient behaviors typical of the originated AE signals. More specifically, both the continuous and the discrete WT alternatives have been explored to find the best compromise between time-frequency resolution and computational complexity in view of extreme edge deployments. Furthermore, the event-driven capabilities of neuromorphic architectures (and spiking neural networks (SNNs) in particular) in processing spiky and sparse temporal information are exploited to retrieve ToA in a beyond state-of-the-art power-efficient manner and negligible loss of performance with respect to standard models. Therefore, we aim at combining the superior performances in ToA identification enabled by the WT operator with the unique energy saving disclosed by spiking hardware and software. Experimental tests executed on a metallic plate structure demonstrated that WT combined with SNN can achieve high precision (median values less than 5 cm) in ToA estimation and AE source localization even in the presence of relevant noise (signal-to-noise ratio down to 2 dB), while its deployment on dedicated neuromorphic architectures can reduce by six orders of magnitude the power expenditure per inference when compared to standard convolutional architectures.