A causal learning approach to in-orbit inertial parameter estimation for multi-payload deployers.
- Paper ID
89332
- DOI
- author
- company
Cranfield University, UK; Cranfield University
- country
United Kingdom
- year
2024
- abstract
The problem of parameter estimation in dynamical systems is ever present in the field of space applications, where control system robustness is required for day-to-day operations. Traditional techniques for on-line state estimation are typically used to this end, such as system observers that work with an extended state representation that includes system parameters. In the special case of payload deployers, the inertial parameters of the carrier spacecraft will change after each deployment event. Thus, a recalibration of attitude control systems that takes into consideration the new moments of inertia is required for optimal controller performance. In this work, a machine learning based estimator is proposed for the payload deployer case, that takes advantage of the a priori expected changes of inertial parameters. In contrast to traditional observer-based approaches which require knowledge of the system’s dynamics, the proposed methodology can be applied to measured or experimental data without any prior knowledge of the system. The methodology proposed herein is based on causal learning, i.e. learning from the responses under actuation of different spacecraft configurations (inertial parameter sets), in order to produce an optimised classifier that can be used to distinguish between them. The actuation is comprised of finite sequences of thruster firings, impacting a generalised torque on the carrier spacecraft that is reproducible in any controllable system. Simulation of attitude dynamics for the possible sets of inertial parameters and input sequences is used for training data generation, and the robustness of the proposed method is increased by injecting sensor noise and system disturbances during the training phase. By training on the system’s responses across multiple input sequences and then applying measures of time-series similarity and F1-score an optimal actuation sequence can be chosen, either for one specific system configuration or for the overall set of possible configurations. This allows for both estimation of the inertial parameter set without any prior knowledge of state, as well as validation of transitions between different configurations. Results show that it is possible for more than one actuation sequence to yield highly accurate (F1 $\ge$ 0.95) distinction capabilities, and unlike traditional methods this allows to consider the actuation-identification interaction as a multi-objective optimisation. Additional objectives that are included in this work are minimisation of fuel expenditure, and divergence of the final spacecraft state from the starting (rest) state.