PIAN: A physics-informed assimilation neural network for temporal super-resolution reconstruction of sensor data in satellite attitude control system

High-resolution sensor data is essential for attitude adjustment, orbit control, and fault diagnosis of satellite attitude control systems. Nevertheless, sensor data from these systems exhibit obvious pseudo-periodicity, randomness, and irregular sampling characteristics, which bring significant challenges to temporal super-resolution reconstruction. To address these, this paper proposes a temporal super-resolution reconstruction method based on a physics-informed assimilation neural network. First, a neural ordinary differential equation model based on an adaptive reconstruction window is employed to mitigate the impact of irregular data sampling on modelling. Secondly, a novel super-resolution framework that combines data-driven and physics-driven methods is proposed to enhance reconstruction accuracy. Among them, the data-driven component utilizes a multi-layer stacked convolution-deconvolution network to reconstruct the transient random element of the sensor data. Simultaneously, the physics-driven component devises an innovative method for control path generation and incorporates neural control differential equations to model the continuous operational process of the system, thus facilitating super-resolution reconstruction of the pseudo-periodic trend component within the data. Subsequently, the particle filter is employed to integrate the results above, thereby mitigating the effects of data randomness and pseudo-periodicity on reconstruction. Finally, within a maximum a posteriori framework, the conservation of angular momentum functions as a physical constraint to improve data consistency and physical validity during model training. Comparative experiments on satellite in-orbit operation datasets demonstrate that the proposed method outperforms other state-of-the-art methods under specific super-resolution factors, effectively improving the resolution of irregularly sampled data. This provides support for spacecraft status assessment and link resource optimization.