Understanding what an aircraft is about to do next like climb, descend, cruise, or fail is key for building safer and smarter flight systems. Most models today use deep learning to analyze engine sensor data, but ignore physical constraints, making them less reliable and poorly generalizable across different flight conditions. We propose AeroSenseNet, a novel Transformer-based model that predicts aircraft maneuver intent from raw turbofan engine sensor sequences while adhering to basic Newtonian motion laws. The model incorporates a physics-informed loss term based on estimated velocity change ( \( \dot{v} \) ), guiding it toward physically consistent decisions. Trained on the NASA C-MAPSS FD001 dataset, AeroSenseNet achieves 96.34% accuracy and a macro F1-score of 0.94, significantly outperforming a baseline model without physics regularization. Attention visualizations show interpretable patterns aligned with maneuver transitions. A zero-shot generalization test on the FD002 dataset further highlights the challenge of varying flight regimes and the value of physics-aware learning. AeroSenseNet is the first framework to combine Transformers and flight dynamics for interpretable aircraft intent prediction from sensor time series. The code and data preprocessing pipeline are available at: https://github.com/satyamcser/AeroSenseNet .