Sequence-to-Sequence Models for 4D Trajectory Prediction of Tower Crane Loads

During tower crane operations, hazards from loads striking objects and powerlines pose significant safety risks, leading to severe consequences such as equipment damage, project delays, injuries, and even fatality. Existing crane control methods often overlook the dynamic interplay of operator behavior, crane kinematics, and site conditions, creating a gap in robust, data-driven collision avoidance solutions. To address this issue, this study introduces a 4D trajectory prediction framework powered by advanced machine learning models to predict load movements and enhance collision avoidance. A tower crane operation simulation system was developed to capture the behavior of novice operators. The simulator integrates synthesized signals within a Unity-based engine to collect real-time sensor data to track load trajectories. Deep learning models, specifically sequence-to-sequence (Seq2Seq) models with various attention mechanisms, were employed to predict load trajectories over time. Results demonstrate that the framework accurately predicts future load positions and provides timely warnings for potential collisions. The framework's adaptability also facilitates integration into various crane operation scenarios to reduce risk and improve operational efficiency.