• The traditional acoustic tap test is augmented by an end-to-end TinyML framework. • A lightweight CNN is deployed on a smartphone for real-time, on-device signal processing. • The model is trained and validated on a unique dataset collected from in-service aircraft. • High accuracy (F1-Score 0.96) is demonstrated on dataset samples collected in field conditions. Assessing the structural integrity of aerospace composites remains a significant challenge for conventional Non-Destructive Testing techniques. These methods are frequently hindered by high operational costs, time-consuming procedures, and a reliance on controlled environments. This paper introduces an end-to-end framework that leverages Tiny Machine Learning for the real-time, on-device detection of subsurface anomalies in composite structures. Unlike prior methods relying on computationally intensive feature extraction, our proposed Squeeze and Excitation Convolutional Neural Network (SE-CNN) processes raw acoustic waveforms directly on a resource constrained smartphone. The model was trained and evaluated on a comprehensive dataset comprising in situ recordings from operational Unmanned Aerial Vehicles and a custom fabricated calibration panel. The system demonstrates effective detection capabilities with an F1-Score of 0.96 on real-world data, while maintaining an inference latency of just 1.03 ms and a minimal model footprint of 22 KB. This research validates the successful deployment of deep learning on the edge, presenting a practical and scalable solution that transforms the subjective manual tap test into an objective, data-driven process, thereby providing field engineers with a standardized tool to minimize inspection variability and human error.
Kfir et al. (Sun,) studied this question.