Kaixin Cui
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Kaixin Cui has a total of 19 co-authors. Here are the top co-authors by number of shared publications:
- Dawei Shi
Beijing Institute of Technology
9 shared publications - Zhigang Liu
Beijing Institute of Spacecraft System Engineering, Beijing, China
5 shared publications - Haijin Li
Beijing Institute of Spacecraft System Engineering, Beijing, China
3 shared publications - Junzheng Wang
School of Automation, Beijing Institute of Technology, Beijing, China
3 shared publications - Dong Yang
Beijing Institute of Spacecraft System Engineering, Beijing, China
2 shared publications Hanjing Fu
MIIT Key Laboratory of Servo Motion System Drive and Control, School of Automation, Beijing Institute of Technology, Beijing, China
2 shared publications
Latest Publications
- 3 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-11
Physics-Enhanced Transfer Learning for the State-of-Charge Estimation of Spacecraft Li-Ion Batteries Using Bi-LSTM Networks
In this work, the problem of state-of-charge (SOC) estimation for batteries with different chemical properties and different operating conditions is investigated based on a physics-enhanced bidirectional long short-term memory (PEBi-LSTM) network. The first physical layer of the PEBi-LSTM network learns the open-circuit and polarization voltages of batteries under various operating conditions by identifying a first-order RC equivalent circuit model, simplifying the mapping between measurement variables and SOC. Then, a two-stage transfer strategy is proposed to estimate the SOC of different lithium batteries. In the first stage, limited labeled data are used to fine-tune the Bi-LSTM layers to extract the effective features of new batteries and ensure model stability. Next, the mapping relationships in the fully connected layers are optimized by the adversarial transfer method, which reduces the data distribution differences between new and original batteries. The effectiveness of the PEBi-LSTM algorithm is verified by comparative experiments on the public dataset. In addition, experiments on a real spacecraft lithium battery platform show that the designed two-stage transfer method reduces errors at -20°C by 74.25% and 45.6% compared with the fine-tuning transfer method and the adversarial transfer method, respectively.
- 3 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-11
Physics-Enhanced Transfer Learning for the State-of-Charge Estimation of Spacecraft Li-Ion Batteries Using Bi-LSTM Networks
In this work, the problem of state-of-charge (SOC) estimation for batteries with different chemical properties and different operating conditions is investigated based on a physics-enhanced bidirectional long short-term memory (PEBi-LSTM) network. The first physical layer of the PEBi-LSTM network learns the open-circuit and polarization voltages of batteries under various operating conditions by identifying a first-order RC equivalent circuit model, simplifying the mapping between measurement variables and SOC. Then, a two-stage transfer strategy is proposed to estimate the SOC of different lithium batteries. In the first stage, limited labeled data are used to fine-tune the Bi-LSTM layers to extract the effective features of new batteries and ensure model stability. Next, the mapping relationships in the fully connected layers are optimized by the adversarial transfer method, which reduces the data distribution differences between new and original batteries. The effectiveness of the PEBi-LSTM algorithm is verified by comparative experiments on the public dataset. In addition, experiments on a real spacecraft lithium battery platform show that the designed two-stage transfer method reduces errors at -20°C by 74.25% and 45.6% compared with the fine-tuning transfer method and the adversarial transfer method, respectively.
- 17 July 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 296
Quantization-aware distributed deep reinforcement learning for dynamic multi-robot scheduling
- 17 July 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 296
Quantization-aware distributed deep reinforcement learning for dynamic multi-robot scheduling
- 2 May 2025
- Published by Elsevier in Computers & Operations Research
- Vol. 182
Meta-learning for dynamic multi-robot task scheduling
- 2 May 2025
- Published by Elsevier in Computers & Operations Research
- Vol. 182
Meta-learning for dynamic multi-robot task scheduling
- 18 February 2025
- Published by Elsevier in Control Engineering Practice
- Vol. 158
Dynamic alarm monitoring with data-driven ellipsoidal threshold learning
- 18 February 2025
- Published by Elsevier in Control Engineering Practice
- Vol. 158
Dynamic alarm monitoring with data-driven ellipsoidal threshold learning
- 12 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Aerospace and Electronic Systems
- Vol. 61 (3) , 7597-7609
Domain Similarity Meta-Learning for Lithium-Ion Battery State-of-Health Estimation of Spacecraft Systems
Data limitation caused by the difficulty of data acquisition, the high cost of data collection, and the changes of working conditions is a serious obstacle to the accurate lithium-ion battery SOH estimation for spacecrafts system. To achieve accurate few-shot lithium-ion battery state of health(SOH) estimation and enhance the adaptability of learning methods across varying working conditions, a domain similarity model agnostic meta-learning method (DS-MAML) is proposed. First, we design a dynamic-static feature extraction method to adequately exploit information from the limited battery data and link the obtained features by gray correlation coefficients to collect sufficient information. Then, by calculating the domain similarity between the training tasks through the maximum mean discrepancy algorithm, the training tasks are ranked to reduce the re-training time. Finally, an LSTM model is added to the meta-learning framework to capture the long-term dependency relationships between SOH and time series of the voltage, and the ranked tasks are embedded in the meta-training process to improve adaptability in different working conditions. The effectiveness of the proposed method is validated by NASA and MIT datasets, and comparative experimental results illustrate that the proposed method average error is reduced by 73%, and the running speed is increased by 20% compared with traditional MAML in few-shot scenarios.
- 12 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Aerospace and Electronic Systems
- Vol. 61 (3) , 7597-7609
Domain Similarity Meta-Learning for Lithium-Ion Battery State-of-Health Estimation of Spacecraft Systems
Data limitation caused by the difficulty of data acquisition, the high cost of data collection, and the changes of working conditions is a serious obstacle to the accurate lithium-ion battery SOH estimation for spacecrafts system. To achieve accurate few-shot lithium-ion battery state of health(SOH) estimation and enhance the adaptability of learning methods across varying working conditions, a domain similarity model agnostic meta-learning method (DS-MAML) is proposed. First, we design a dynamic-static feature extraction method to adequately exploit information from the limited battery data and link the obtained features by gray correlation coefficients to collect sufficient information. Then, by calculating the domain similarity between the training tasks through the maximum mean discrepancy algorithm, the training tasks are ranked to reduce the re-training time. Finally, an LSTM model is added to the meta-learning framework to capture the long-term dependency relationships between SOH and time series of the voltage, and the ranked tasks are embedded in the meta-training process to improve adaptability in different working conditions. The effectiveness of the proposed method is validated by NASA and MIT datasets, and comparative experimental results illustrate that the proposed method average error is reduced by 73%, and the running speed is increased by 20% compared with traditional MAML in few-shot scenarios.
- 12 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 20 (12) , 14401-14410
A Physics-Informed Event-Triggered Learning Approach to Long-Term Spacecraft Li-Ion Battery State-of-Charge Estimation
In this work, an event-triggered learning problem for state-of-charge (SOC) estimation of long-term spacecraft li-ion batteries with system aging and disturbances is investigated based on a physics-informed long short-term memory (PI-LSTM) network. An equivalent circuit model and a pretrained Gaussian process regression model are integrated into a long short-term memory (LSTM) network, which is trained and updated quickly with limited transmission data. By considering noisy data and physical constraints simultaneously, the PI-LSTM approach provides interpretable dynamic models for the long-term battery SOC estimation. Then, an unscented Kalman filter is proposed to estimate the SOC performance. By using weighted average voltage prediction errors, an event-triggering condition is established to guarantee the estimation performance with a reduced signal transmission rate. The effectiveness of the proposed approach is validated through experiments on a real spacecraft Li-ion battery platform, which achieves the SOC estimation error of less than 2%, and the maximum voltage prediction error is reduced by 61% after updating the PI-LSTM model.
- 12 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 20 (12) , 14401-14410
A Physics-Informed Event-Triggered Learning Approach to Long-Term Spacecraft Li-Ion Battery State-of-Charge Estimation
In this work, an event-triggered learning problem for state-of-charge (SOC) estimation of long-term spacecraft li-ion batteries with system aging and disturbances is investigated based on a physics-informed long short-term memory (PI-LSTM) network. An equivalent circuit model and a pretrained Gaussian process regression model are integrated into a long short-term memory (LSTM) network, which is trained and updated quickly with limited transmission data. By considering noisy data and physical constraints simultaneously, the PI-LSTM approach provides interpretable dynamic models for the long-term battery SOC estimation. Then, an unscented Kalman filter is proposed to estimate the SOC performance. By using weighted average voltage prediction errors, an event-triggering condition is established to guarantee the estimation performance with a reduced signal transmission rate. The effectiveness of the proposed approach is validated through experiments on a real spacecraft Li-ion battery platform, which achieves the SOC estimation error of less than 2%, and the maximum voltage prediction error is reduced by 61% after updating the PI-LSTM model.
- 17 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ASME Transactions on Mechatronics
- Vol. 29 (5) , 3546-3555
Physics-Informed Neural Network for Spacecraft Lithium-Ion Battery Modeling and Health Diagnosis
The modeling of lithium-ion batteries is essential for the battery management. In space operating conditions, no additional tests can be performed, and the battery is usually not discharged to empty, resulting in less data available for analysis. Furthermore, the battery characteristics are influenced by factors, such as service time and temperature, so the model should be able to be updated with limited data. This article investigates a novel physics-informed neural network (PINN) model, which integrates the equivalent-circuit model (ECM) into the recurrent neural network (RNN) framework to reconstruct RNN structure. The integration of ECM adds physical constraints to the training of RNN, so that the model conforms to both the physical knowledge and data distribution, which reflects the impact of environment and aging. The approach is validated through simulation and experiments based on a real spacecraft lithium-ion battery cell. Results indicate that the proposed model outperforms a white-box model and an enhanced ECM on real-time modeling and discharge profiles prediction. Moreover, the battery state of health is estimated based on the parameters identified by this method, with an error of less than 2%.
- 17 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ASME Transactions on Mechatronics
- Vol. 29 (5) , 3546-3555
Physics-Informed Neural Network for Spacecraft Lithium-Ion Battery Modeling and Health Diagnosis
The modeling of lithium-ion batteries is essential for the battery management. In space operating conditions, no additional tests can be performed, and the battery is usually not discharged to empty, resulting in less data available for analysis. Furthermore, the battery characteristics are influenced by factors, such as service time and temperature, so the model should be able to be updated with limited data. This article investigates a novel physics-informed neural network (PINN) model, which integrates the equivalent-circuit model (ECM) into the recurrent neural network (RNN) framework to reconstruct RNN structure. The integration of ECM adds physical constraints to the training of RNN, so that the model conforms to both the physical knowledge and data distribution, which reflects the impact of environment and aging. The approach is validated through simulation and experiments based on a real spacecraft lithium-ion battery cell. Results indicate that the proposed model outperforms a white-box model and an enhanced ECM on real-time modeling and discharge profiles prediction. Moreover, the battery state of health is estimated based on the parameters identified by this method, with an error of less than 2%.
- 21 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Electronics
- Vol. 71 (7) , 7954-7962
Behavioral-Based Monitoring of Networked Systems With Application to Spacecraft Li-Ion Battery Health Management
Performance monitoring is essential for the safety maintenance and health management of networked systems. In this work, we propose a behavioral-based monitoring approach for networked systems, which directly utilizes input and output measurements without parametric identification. A regularized cost function and system dynamic constraints are designed to reduce the disturbance of system nonlinearity and noise, which can be used to predict future trajectories of networked systems. Missing data during the long-distance transmission are also estimated through the fundamental lemma in behavioral system theory. Then, a statistical inference strategy based on the difference between the predicted trajectories and the real measurement trajectories is introduced to detect system faults. The effectiveness of the proposed approach is illustrated through experimental applications to the spacecraft Li-ion battery health management, which includes charging and discharging trajectory prediction, missing data estimation, and voltage sensor and overdischarge fault detection.
- 21 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Electronics
- Vol. 71 (7) , 7954-7962
Behavioral-Based Monitoring of Networked Systems With Application to Spacecraft Li-Ion Battery Health Management
Performance monitoring is essential for the safety maintenance and health management of networked systems. In this work, we propose a behavioral-based monitoring approach for networked systems, which directly utilizes input and output measurements without parametric identification. A regularized cost function and system dynamic constraints are designed to reduce the disturbance of system nonlinearity and noise, which can be used to predict future trajectories of networked systems. Missing data during the long-distance transmission are also estimated through the fundamental lemma in behavioral system theory. Then, a statistical inference strategy based on the difference between the predicted trajectories and the real measurement trajectories is introduced to detect system faults. The effectiveness of the proposed approach is illustrated through experimental applications to the spacecraft Li-ion battery health management, which includes charging and discharging trajectory prediction, missing data estimation, and voltage sensor and overdischarge fault detection.
- 23 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Aerospace and Electronic Systems
- Vol. 59 (4) , 3766-3777
Event-Triggered Deep Reinforcement Learning for Dynamic Task Scheduling in Multisatellite Resource Allocation
In this work, we investigate the problem of multi-satellite resource allocation for expected long-term performance optimization with a dynamic task network model, where communication tasks generated by task satellites are expected to be transmitted by resource satellites in the application layer, and the set of tasks changes with satellite orbital motions. The features of the tasks include priority, execution duration, visible time, etc. Since the feature information has a high dimension and changes with time, the scheduling problem is formulated as a dynamic combinatorial optimization problem and a receding-horizon task scheduling algorithm based on the event-triggered deep reinforcement learning is proposed. A residual-fully connected network is designed to extract the features of the complex task network model, and a deep double Q-learning iteration with the experience replay memory mechanism is employed to change the allocation strategy by evaluated rewards adaptively. An event-triggered strategy is then proposed to handle urgent tasks online. Numerical simulations show the performance improvement of the proposed algorithm. For the scenario of 50 task satellites and 10 resource satellites, the proposed algorithm achieves 4.1%, 5.9%, and 11.4% higher reward scores than the static deep reinforcement learning algorithm, the data-driven parallel scheduling algorithm, and the improved genetic algorithm, respectively. The computation time of the proposed algorithm is only 34.7% and 21.3% of that of the latter two algorithms, and is similar to that of the static deep reinforcement learning algorithm.
- 23 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Aerospace and Electronic Systems
- Vol. 59 (4) , 3766-3777
Event-Triggered Deep Reinforcement Learning for Dynamic Task Scheduling in Multisatellite Resource Allocation
In this work, we investigate the problem of multi-satellite resource allocation for expected long-term performance optimization with a dynamic task network model, where communication tasks generated by task satellites are expected to be transmitted by resource satellites in the application layer, and the set of tasks changes with satellite orbital motions. The features of the tasks include priority, execution duration, visible time, etc. Since the feature information has a high dimension and changes with time, the scheduling problem is formulated as a dynamic combinatorial optimization problem and a receding-horizon task scheduling algorithm based on the event-triggered deep reinforcement learning is proposed. A residual-fully connected network is designed to extract the features of the complex task network model, and a deep double Q-learning iteration with the experience replay memory mechanism is employed to change the allocation strategy by evaluated rewards adaptively. An event-triggered strategy is then proposed to handle urgent tasks online. Numerical simulations show the performance improvement of the proposed algorithm. For the scenario of 50 task satellites and 10 resource satellites, the proposed algorithm achieves 4.1%, 5.9%, and 11.4% higher reward scores than the static deep reinforcement learning algorithm, the data-driven parallel scheduling algorithm, and the improved genetic algorithm, respectively. The computation time of the proposed algorithm is only 34.7% and 21.3% of that of the latter two algorithms, and is similar to that of the static deep reinforcement learning algorithm.