Rare-event detection of incipient sediment motion from smart-particle time series using deep learning


Incipient sediment motion in turbulent flows remains difficult to characterize and predict because the underlying hydrodynamic forces are highly intermittent and events are sparse in time, even in well-controlled experiments. This study investigates whether temporal deep-learning architectures can detect the onset of particle motion directly from high-frequency velocity time series measured by an instrumented “smart sphere” [1, 2], without explicit force or torque measurements. The workflow includes detrending and cleaning of raw signals, physics-informed signal transforms (e.g. smoothed velocity, acceleration, jerk, and kinematic impulse proxies), segmentation with sliding windows, and supervised training of temporal deep-learning architectures, including recurrent, convolutional, and attention-based models, using class-imbalance mitigation such as focal loss, class weighting, and data augmentation.
Hyperparameter optimization is performed automatically with Optuna, and model performance is assessed using ROC and precision–recall curves, confusion matrices and time-resolved prediction performance. Results show that all tested architectures can learn consistent kinematic signatures preceding incipient motion from single-axis velocity time series, with models incorporating attention mechanisms achieving the highest recall on rare motion-onset events, consistent with their ability to focus on intermittent, high-magnitude kinematic bursts preceding entrainment. These findings demonstrate that deep learning applied to smart-particle sensor data can provide an efficient, non-intrusive tool for particle-scale sediment transport monitoring and real-time–capable event detection. The approach is directly relevant to the session’s focus on particle-scale transport mechanics and data-driven upscaling, and opens avenues for integrating deep-learning-based event detection into multi-scale sediment transport models in geophysical and engineered flows.

References
[1] Al-Obaidi, K., Xu, Y., & Valyrakis, M. (2020). The design and calibration of instrumented particles for assessing water infrastructure hazards. Journal of Sensor and Actuator Networks, 9(3), 36.
[2] AlObaidi, K., & Valyrakis, M. (2021). Linking the explicit probability of entrainment of instrumented particles to flow hydrodynamics. Earth Surface Processes and Landforms, 46(12), 2448-2465.