Adaptive importance sampling-enhanced B-spline PINN for hypersonic vehicle competency assessment with faults and uncertainties

Competency assessment of hypersonic vehicles, which evaluates real-time task capability under dynamic conditions, is critical for decision-making. The existing methods struggle with the complex, nonlinear dynamic and multi-constraint tasks involving faults and uncertainties encountered by hypersonic vehicles. To tackle these challenges, this study proposes a novel online task competency assessment method that accounts for faults and multi-source uncertainties, and explores vehicle competency upon fault occurrences. A B-spline physics-informed neural network was developed to replace the computationally expensive trajectory simulation, directly learning the mapping from the trajectory before faults, along with fault and deviation parameters, to the trajectory after faults. To ensure the accuracy and efficiency of trajectory prediction, the BS-PINN incorporates a B-spline layer for representing time-series trajectory data with inconsistent lengths, dual prediction heads for extracting both global and local trajectory features, and a computationally efficient physics-informed loss function to enhance convergence during training. Additionally, an adaptive importance sampling method is proposed to significantly enhance the real-time performance of task competency evaluation in conjunction with BS-PINN. The effectiveness and advantages of the proposed method are verified through ablation and validation tests involving a hypersonic vehicle with various complex tasks on computer and embedded platforms.