Collins Aerospace project for machine understanding of human behavior
- Injected the description of abnormal behaviors from human-domain expert knowledge into learning methods via the introduction of parametric Signal Temporal Logic (pSTL) formulas.
- Implemented a deep learning algorithm to obtain learned STL formula for describing or classifying different human behaviors.
- Solved human behavior classification problems via the learned STL criterion from deep neural network models with the IMU data from human behaviors.
Reachability analysis for falsification of Baron58 airplane linearized system
- Used Hamilton-Jacobi-Bellman (HJB) level set method to compute backward reachable sets, which is used for safety verification and falsification, for the identified system.
- Compared various reachability analysis methods such as the HJB method and the zonotope-based method, and study the differences between them.
- Implemented sequential linearization and splitting procedure to extend the zonotope-based reachability analysis method, which is used for linear systems, to nonlinear systems.
- Implemented sequential linearization and splitting procedure to extend the zonotope-based reachability analysis method to constrained-zonotope-based reachability analysis.