Moving target trajectory prediction based on Dropout-LSTM and Bayesian inference for long-time multi-satellite observation

Multi-satellite cooperation facilitates long-time and wide-area observation of moving targets and has received considerable attention. An accurate and real-time trajectory prediction of an observed target is the primary premise of multi-satellite cooperative observation that attempts to address the field-of-view switching problem for different satellites. However, the existing methods do not consider the non-linear and time-variant satellite observation process, and the interpretability requires improvement. Therefore, a trajectory prediction model is constructed with decoupled linear Gaussian and non-linear Gaussian parts, based on the analysis of the satellite observation process and the verification of the normal distribution of the observation error. Thereafter, a trajectory prediction method, comprising long short-term memory networks and Bayesian inference, is proposed. The proposed method not only enables the long short-term memory with dropout variational inference to automatically extract features and learn more complex spatiotemporal patterns, but also employs Bayesian inference to introduce the satellite orbit and attitude information as known variables and improve the spatiotemporal smoothness. Finally, a simulation system of moving target trajectory prediction is established, and the results indicate that our approach is superior in terms of the prediction accuracy and generalization properties, compared with the existing methods, particularly when the observed target is manoeuvring rapidly. Consequently, the proposed approach provides a technical basis for achieving long-time multi-satellite observation of the moving target.

Automated summary

The research proposes a trajectory prediction model for multi-satellite cooperative observation of moving targets, consisting of linear Gaussian and non-linear Gaussian parts, as well as a prediction method combining long short-term memory networks and Bayesian inference.