A model-driven dual-derivation framework for quantitative fault detection in satellite power system

Satellite power system (SPS) fault detection is of great significance to ensure the safety and stability of satellites. On-orbit SPS can divide 11 near-mutation operating conditions (OCs) in 4 types of sunlight regions. Combined with limited fault samples and high-dimensional, coupled, and noisy telemetry data, the accuracy of data- or knowledge-driven SPS fault detection is poor. Therefore, this work first comprehensively considers the mechanism model of SPS and the quantitative analysis results of corresponding faults, based on which an SPS fault behavior model is configured. By combining specific driving and parameter updating methods, strong support is provided for on-orbit SPS digital twin and fault detection. Then, a model-driven dual-derivation quantitative fault detection framework that combines accuracy and robustness is proposed. To be specific, an adaptive integral residual (AIR) algorithm for constructing SPS OCs is developed, which combines telemetry data with twin data to determine fault states and obtain fault information. Using the tree-structured Parzen estimator (TPE), iteratively adjust the model's failure modes and parameters to obtain simulated data for the current fault. By comparing it with fault telemetry data, determine whether the current failure modes and parameters meet the requirements of quantitative fault detection. Finally, a semi-physical experimental platform was established, and experimental results confirmed the framework's capability to accurately differentiate between different levels of faults. Specifically, the quantitative detection accuracy for typical faults reached 100%. Additionally, we designed seven accuracy and robustness indicators, all of which yielded optimal results when compared with common methods. Through experimental analysis of search space optimization methods, the universality of optimization methods has been demonstrated.