Data-Driven Identification of Cavitational Behavior During Vibrodriving

During vibrodriving, pile penetration dynamics can be broadly divided into non-cavitational and cavitational regimes. In the cavitational regime, excessive acceleration may cause intermittent pile--soil detachment, thereby affecting energy transfer, soil disturbance, and final bearing capacity. Identifying the prevailing driving regime is essential for adaptive control and for assessing installation efficiency. However, conventional measurements of reaction force and displacement are often impractical in field applications because of sensor cost and integration constraints. This study proposes a data-driven classification approach to distinguish between cavitational and non-cavitational vibrodriving regimes. Laboratory model tests are performed using a scaled vibrodriver prototype with independently controlled eccentric masses. Reaction forces are measured using a piezoelectric load cell mounted at the pile head, tip pressure is measured beneath the pile toe, and vertical displacement is obtained from an encoder. The reference measurements enable cycle-wise classification using two independent methods. Both methods show consistent results and capture the transition zone between the two regimes. Since the reference measurements are often too costly or impractical for real-time field applications, labeled acceleration signals derived from force, pressure, and displacement measurements are used as input for a machine learning model. A supervised learning workflow based on a convolutional neural network is trained for classification. The results show that the network achieves a test accuracy of 99.94\% and correctly predicts nearly all cycles. The accelerometer-based classification provides a promising low-cost basis for monitoring and real-time control of vibrodriving systems.