Stringent confidentiality policies in aerospace domains severely limit accessible spacecraft observation data. However, given the substantial costs of space missions and the proliferation of global satellite constellations, advancing operational maintenance technologies for aerospace equipment is imperative. Consequently, we publicly release the XJTU-SPS dataset to facilitate research in this field.
XJTU-SPS targets spacecraft power system (SPS), one of the most failure-prone and critically consequential subsystems in spacecraft.
As a representative complex industrial system with multi-sensor network, spacecraft power systems exhibit intricate physical couplings and diverse operational regimes. This dataset can be utilized in other domains for validating algorithm generalizability.
Specifically, XJTU-SPS consists of four sub-datasets: (1) XJTU-SPS for Forecasting or Reconstruction (XJTU-SPS for F or R), (2) XJTU-SPS for Work Mode Recognition (XJTU-SPS for MR), (3) XJTU-SPS for Anomaly Detection (XJTU-SPS for AD), and (4) XJTU-SPS for Fault Localization / Fault Diagnosis (XJTU-SPS for FL or FD).
Meanwhile, XJTU-SPS is supplemented with fully aligned mathematical-physical models enabling simulations and physical knowledge provision, thereby we also anticipate that XJTU-SPS can provide some support to researchers in Physics-Informed Neural Networks (PINNs), digital twins, and predictive control.
Keywords: spacecraft, power system, PINN, digital twin, monitoring and operation maintenance, time series forecasting, time series reconstruction, work mode recognition, anomaly detection, fault localization, fault diagnosis.
The SPS comprises the following subsystems: solar array (SA), battery sets (BAT), battery charge regulator (BCR), battery discharge regulator (BDR), shunt regulator (SR), power distribution module (PDM), loads (Loads), and others. The SA is the power source of the SPS, responsible for converting solar radiation into electrical energy, and its output power is affected by factors such as solar radiation, temperature, and shadow. The BAT is the power storage unit of the SPS, responsible for storing the electrical energy provided by the SA. The BCR is the charge control unit of the SPS, responsible for controlling the charging of the BAT. The BDR is the discharge control unit of the SPS, responsible for controlling the discharge of the BAT to the Loads. The SR is the energy regulation unit of the SPS, it works with the BCR and BDR to complete the following functions: • When located in the sunlight area and the sunlight is sufficient, the SA charges the BAT and supplies power to the Loads, • When located in the sunlight area and the sunlight is insufficient, the SA and BAT jointly supply power to the Loads, • When located in the shadow area, the SA does not work, and the BAT supplies power to the Loads. The PDM is the power distribution unit of the SPS, responsible for distributing electrical energy to various Loads. The Loads are the power consumption unit of the spacecraft power system, and their power requirements are affected by factors such as the working state of the spacecraft and mission requirements.
Note: Although we provide the corresponding physical simulation model, our fault data is not generated through the physical simulation model. This experimental platform is not a semi-physical simulation platform, but a physical hardware fault injection simulation platform.
`charge' means SA only charges BAT without any load, `shunt' means SA provides energy for both LOAD and BAT, `joint' means SA and BAT provide energy for LOAD simultaneously, `leisure' means there is no load and BAT is not charging, and `discharge' means only BAT supplies energy to LOAD.
Specifically, the dataset simulates the on-orbit operational environment of spacecraft in low Earth orbit (LEO), which is the most common type of spacecraft orbit, where many important spacecraft such as communication satellite constellations and space stations are deployed, and their power systems frequently switch between operational conditions, making anomalies and faults representative and typical. Its single orbital cycle is set to 95 minutes, with the spacecraft located in the sunlit area for the first 60 minutes and in the shadow area for the remaining 35 minutes. Under normal circumstances, to maximize power generation, the spacecraft operates in sun orientation mode. However, when performing tasks such as earth observation imaging and radar alignment with ground stations, it needs to switch to ground orientation mode.
Details: (1) 0-10min: The spacecraft enters the sunlit area, no task is performed, sunlight is sufficient, the SA charges the BAT, and it undergoes a very short constant current (CC) charging stage followed by a rapid transition to constant voltage (CV) charging stage, in fast charging mode; (2) 10-25min: Non-orientation Task 1 is performed, sunlight is sufficient, the SA's energy is adequate to simultaneously power the load and charge the BAT, and there is even excess power being shunted (or MPPT working at non-maximum power point), in shunt mode; (3) 25-35min: Task 1 stops, sunlight is sufficient, the SA charges the BAT, and later only compensates for the self-discharge loss of the BAT, in trickle charging mode; (4) 35-50min: Due to tasks such as earth observation imaging or radar alignment with ground stations, it switches to ground orientation mode, sunlight is insufficient, the SA cannot independently power the load, the BAT and SA jointly supply power to the load, in joint power supply mode; (5) 50-60min: Task 2 stops, sunlight is sufficient, the SA charges the BAT, in fast charging mode; (6) 60-70min: The spacecraft enters the shadow area where there is no solar irradiation. The orientation mode depends on energy remaining and task requirements. No task is performed at this time, in idle mode; (7) 70-85min: Non-orientation Task 3 is performed. Due to no irradiation, the SA cannot work, and the BAT supplies power to the load, in discharge mode; (8) 85-95min: Task 3 stops and returns to idle mode. (9) After 95min, this cycle ends and returns to the first step, starting the next cycle.
We did not measure all battery groups, but only the 2nd, 3rd, and 4th groups of batteries, in fact, the other 2 groups of batteries exhibit similar observational properties.
XJTU-SPS for F or R does not contain any anomalies or faults. This dataset is used for time series forecasting or reconstruction tasks, typically involving spacecraft state forecasting, parameter trend extrapolation, missing data reconstruction, mission planning, digital twin construction, and predictive health trend analysis. The forecasting and reconstruction of spacecraft operational states also form the basis for subsequent anomaly detection and fault diagnosis tasks.
XJTU-SPS for F or R consists of 6 subsets, namely: XJTU-SPS for F or R_4cycles, XJTU-SPS for F or R_18cycles, XJTU-SPS for F or R_24cycles, XJTU-SPS for F or R_34cycles, XJTU-SPS for F or R_90cycles, and XJTU-SPS for F or R_94cycles, which simulate 4, 18, 24, 34, 90, and 94 orbital cycles of spacecraft respectively. Users can select, sample, and truncate the data according to their engineering and project needs.
Visualization example of XJTU-SPS for F or R_18cycles:
XJTU-SPS for work Mode Recognition (XJTU-SPS for MR) do not contain any anomalies or faults. This dataset is used for mode recognition, typically involving energy management, mode switching monitoring, mission phase discrimination and planning, environmental state determination, and payload working state recognition of spacecraft. For complex industrial systems with multi-operational condition switching, performing work condition recognition as the initial stage could effectively reduce the false alarm rates in subsequent anomaly detection and the misdiagnosis rates in fault diagnosis.
XJTU-SPS for MR consists of XJTU-SPS for MR_4cycles, XJTU-SPS for MR_18cycles, XJTU-SPS for MR_24cycles, XJTU-SPS for MR_34cycles, XJTU-SPS for MR_90cycles, and XJTU-SPS for MR_94cycles, which simulate 4, 18, 24, 34, 90, and 94 orbital cycles of spacecraft respectively.
Its mode label information is as follows: 1: Fast Charging, Sunlit Area, No Task; 2: Shunt, Sunlit Area, Non-orientation Task 1; 3: Trickle Charging, Sunlit Area, No Task; 4: Joint Power Supply, Sunlit Area, Ground-orientation Task 2; 5: Idle, Shadow Area, No Task; 6: Discharge, Shadow Area, Non-orientation Task 3. Specifically, (1) Label 1 corresponds to the fast charging condition. Characteristics: the SA output power is large, the BAT experiences a very short constant current (CC) charging followed by an immediate transition to constant voltage (CV) charging state, and the LOAD power is small; (2) Label 2 corresponds to the shunt condition. Characteristics: the SA output power capability is greater than the sum of LOAD and BAT charging demand, the shunt regulator shunts a large current or the MPPT (Maximum Power Point Tracking) algorithm operates at a non-maximum power point; (3) Label 3 corresponds to the trickle charging condition. Characteristics: the BAT charging current is small, which compensates for the self-discharge loss of the BAT and maintains a full charge state; (4) Label 4 corresponds to the joint power supply condition. Characteristics: the SA output power is large, the BAT discharge current is large, and the LOAD power is large; (5) Label 5 corresponds to the idle condition. Characteristics: no power output from SA, small BAT charge/discharge current, and small LOAD power; (6) Label 6 corresponds to the discharge condition. Characteristics: no power output from SA, large BAT discharge current, and large LOAD power. In our published XJTU-SPS dataset, based on the operational characteristics of the power system, we preliminarily classify the conditions into 6 categories. Users can further refine the condition classification according to their actual research needs, such as distinguishing between CC and CV charging, or strictly distinguishing between trickle charging and the late stage of CV charging when it is about to be fully charged based on a strict power thresholds.
Visualization example of XJTU-SPS for MR_4cycles:
XJTU-SPS for Anomaly Detection (XJTU-SPS for AD) is a dataset for spacecraft anomaly detection, which covers a wide range of anomaly situations in spacecraft power systems, effectively evaluating the effectiveness of anomaly detection methods.
XJTU-SPS for AD consists of XJTU-SPS for AD_Train, XJTU-SPS for AD_Test, and XJTU-SPS for AD_Test_AnomalyLabel, which are the training set, test set, and anomaly label file respectively.
Visualization example of XJTU-SPS for AD_Test:
XJTU-SPS for Fault Localization / Fault Diagnosis (XJTU-SPS for FL or FD) is a dataset for spacecraft fault localization or fault diagnosis, which covers a wide range of fault situations in spacecraft power systems, effectively evaluating the effectiveness of fault localization and fault diagnosis methods.
XJTU-SPS for FL or FD consists of normal data files and 17 types of fault data files. Each type of fault is composed of a training file and a test file.
Its label information is as follows: 0: normal; 1: SA_partial component or branch open circuit; 2: SA_partial component or branch short circuit; 3: SA_component deterioration; 4: BCR_open circuit; 5: BCR_short circuit; 6: BCR_increased power loss; 7: BAT_degradation; 8: BAT_open circuit; 9: Bus_open circuit; 10: Bus_short circuit; 11: Bus_insulation breakdown; 12: PDM1_open circuit or short circuit; 13: PDM2_open circuit or short circuit; 14: PDM3_open circuit or short circuit; 15: Load1_open circuit; 16: Load2_open circuit; 17: Load3_open circuit;
Supplementary Note: The exclusion of BAT short-circuit and Load short-circuit fault types stems from experimental safety constraints. While attempting to simulate BAT short-circuits using an external battery pack alongside onboard batteries, we observed significant safety hazards. Furthermore, observations showed that a sudden short circuit immediately triggers the power system's protection mechanisms, producing experimental signatures indistinguishable from those of an open-circuit fault. Therefore, the simulations of these short-circuit faults can be considered as subsumed under the open-circuit fault category.
Visualization example of SA_partial component or branch short circuit:
XJTU-SPS is accompanied by a corresponding mathematical-physical model that enables simulations fully consistent with the dataset. It supports monitoring and maintenance applications and provides comprehensive physical knowledge for researches on digital twins, PINNs, and related topics.
Visualization example of simulation results for several representative channels are presented. For the comprehensive simulation results of all channels, please refer to this study.
If this dataset or model is helpful for your research, you can star our repository and cite our works :
@misc{di2026empoweringallinloophealthmanagement,
title={Empowering All-in-Loop Health Management of Spacecraft Power System in the Mega-Constellation Era via Human-AI Collaboration},
author={Yi Di and Zhibin Zhao and Fujin Wang and Xue Liu and Jiafeng Tang and Jiaxin Ren and Zhi Zhai and Xuefeng Chen},
year={2026},
eprint={2601.12667},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2601.12667},
}
@article{DI2025113380,
title = {PhyGNN: Physics guided graph neural network for complex industrial power system modeling},
author = {Yi Di and Fujin Wang and Zhi Zhai and Zhibin Zhao and Xuefeng Chen},
year = {2025},
journal = {Mechanical Systems and Signal Processing},
volume = {240},
pages = {113380},
issn = {0888-3270},
doi = {https://doi.org/10.1016/j.ymssp.2025.113380},
url = {https://www.sciencedirect.com/science/article/pii/S0888327025010817},
keywords = {Physics guided graph neural network, Spacecraft power system, Multivariate time series, Complex industrial system},
}