Carbon-fiber-reinforced plastics (CFRPs) and similar composites have an improved strength-to-weight ratio over metals, making them ideal for use in demanding lightweight applications such as wind turbines or aircraft. However, these structures are highly susceptible to impacts, I.e., through bird strikes, that can cause internal damage to composites (e.g. delamination) while leaving little to no visual indication on the surface [1]. By applying structural health monitoring (SHM) through sensor enriched smart composites, these events can be detected in real time, reducing the need for regular maintenance and thereby bringing down the operational cost. This is commonly achieved through physics-based impact wave time of arrival (TOA) triangulation, requiring high accuracy sensors. Through the use of machine learning (ML), recent work has also shown low-accuracy, low-cost micro-electromechanical-systems (MEMS) accelerometers to be capable of localizing impacts with sub-millimeter accuracy [2]. As these sensors lack the time resolution required for TOA calculations, the trained convolutional neural networks (CNNs) must identify different physical information encoded in the sensor data that is not known as of yet, making the CNNs impact prediction a black box. In adjacent applications, progress has been made to increase the prediction transparency through explainable artificial intelligence (XAI) [3]. In this study, we use XAI to further the understanding of MEMS-based impact detection systems. To achieve this, we collect real-world impact data on CFRP plates using an automated testbed. To enhance the interpretability of the model’s decision-making process, we apply integrated gradients to detect key patterns within the sensor data, offering insights into the contribution of different features to the impact prediction. Additionally, we use simulations to investigate the underlying significance of various frequencies in the sensor signals, helping to correlate specific frequency ranges with distinct impact characteristics and improve the overall robustness of the detection system. [1] Juhyeong Jang, In Yong Lee, Young-Bin Park, Impact response analysis and Physics-Informed damage classification of sandwich composites using electrical resistance-based self-sensing, Composites Part A: Applied Science and Manufacturing, Volume 190, 2025 [2] A.M. Damm, C. Spitzmüller, A. T. S. Raichle, A .Bühler, P. Weißgraeber, P. Middendorf. Deep learning for impact detection in