Unmanned Aerial Vehicles (UAVs) are widely used in transportation, telecommunications, scientific mapping, and remote sensing. However, their diverse applications make them susceptible to propulsion system faults, particularly in the propellers. To address this concern, signal processing techniques combined with machine learning classifiers have been used to detect damage by analyzing quadrotor vibration data. Despite their effectiveness, these methods often depend on handcrafted features and require high computational resources, posing challenges for real-time embedded implementation. The wavelet scattering transform (WST) provides a robust and interpretable signal representation through cascaded convolution operations. While structurally similar to one-dimensional convolutional neural networks (1D CNNs), WST does not require training or large datasets. Compared to two-dimensional CNNs, 1D CNNs offer lower complexity and shallower architectures, making them more suitable for real-time quadrotor fault detection. However, limited work has explored the trade-offs between WST and 1D CNNs for this application. This study compares WST and 1D CNNs using vibration data in terms of classification performance and execution time. Support vector machines trained on WST coefficients achieve comparable accuracy and F1-score to 1D CNNs, while requiring significantly less computation time, with inference times under five milliseconds compared to over ten milliseconds for CNNs. 2026 The Authors..
