In this paper, we improve the deep deterministic policy gradient algorithm and combine the graph attention network to propose a spacecraft fault diagnosis method. Based on the construction of spacecraft system level and component level knowledge graphs, a unique reward function, policy network and value network are set up according to the structure of spacecraft knowledge graphs and the semantic configuration of reinforcement learning environment. Based on the construction of spacecraft system level and component level knowledge graphs, unique reward functions, environments, policy networks and value networks are set according to the structure and semantics of spacecraft knowledge graphs. We use in orbit data for experimental validation, and the experimental results show that the method can combine systemlevel knowledge graph with component level knowledge graph for hierarchical, fast and accurate fault diagnosis.

With the development of information technology and sensor technology, data driven fault diagnosis technology is one of the key technologies to ensure the efficient and safe operation of large industrial equipment. Due to its powerful feature representation ability and the advantages of feature extraction based on big data, machine learning has become one of the most commonly used feature extraction methods in the field of fault diagnosis. However, the data collected by the monitoring equipment includes a large amount of unlabeled data, and the traditional deep neural network model does not make full use of it, resulting in the waste of some useful information. For unlabeled data, we adopt the idea of unsupervised learning, train a feature extraction model by maximizing mutual information, and on this basis, we design a fault diagnosis method for time series data, and verify it on the public dataset Case Western Reserve University bearing dataset, achieving higher diagnostic accuracy than previous traditional methods. Further verification on satellite monitoring data, our feature extraction model can distinguish different stages of failure and capture the data characteristics of different stages. The results show that the fault diagnosis method based on unsupervised learning proposed in this paper can effectively and fully utilize a large amount of unlabeled data and improve the fault diagnosis accuracy of time series data.

Accurate prediction of the Remaining Useful Life (RUL) of rolling bearings is of paramount importance for ensuring the safe, stable, and reliable operation of engineering equipment. Existing deep learning prediction methods often directly establish a mapping relationship between vibration monitoring data and RUL, typically overlooking the differential states of rolling bearing performance degradation and neglecting the diversity of features extracted by deep learning models, leading to significant bias in RUL prediction results. In light of this, a novel method for dividing the degradation state of rolling bearings and predicting RUL is proposed. Features of bearing vibration signals are extracted, and the Mann Kendall test is employed to judge trends, determining the starting point of the degradation period. The endpoint of the slow degradation period is identified through the trend of normalized singular value correlation coefficients. A rolling bearing RUL prediction model based on a bidirectional long short term memory network with attention (Bi-LSTM-Att) is constructed, and the slow degradation period data and corresponding RUL labels are used to train the prediction model to achieve RUL prediction. The accuracy and effectiveness of the proposed method for bearing RUL prediction are validated through a public bearing dataset.

Aiming at the characteristics of wide sources, diverse forms, large differences and small quantities of spacecraft fault knowledge, and the defects that the fault knowledge contained in telemetry data during spacecraft operation is difficult to be effectively utilized, this paper proposes an ontology entity bidirectional constrained knowledge graph construction method, which adopts a combination of top down and bottom up approaches to construct knowledge graphs and implement graphical fusion of multi source heterogeneous fault knowledge of spacecraft including telemetry data. At the ontology layer, an improved IDEF5 method is proposed to construct the fault knowledge ontology. At the entity layer, three different knowledge extraction methods are proposed to extract knowledge from (semi)structured data (FMEA analysis table, expert rules), unstructured data (fault texts) and telemetry data, and fuse the knowledge in the entity layer, according to the sources of fault data and the degree of structure. The construction of the fault knowledge graph is realized through the bidirectional constraints and collaborative optimization between the ontology layer and the entity layer. In this paper, taking the spacecraft control moment gyroscope as an example, the fault knowledge graph is constructed and displayed visually by using the above method. The feasibility and effectiveness of the method are verified by a case study.

A satellite fault diagnosis approach based on federated learning is proposed to address issues such as single satellite under configuration and incomplete measurement information. Firstly, a fault model for satellites is established, and fault data is generated via the unity simulation environment. Then, a Bidirectional Coordination Network (BicNet) is used to construct local training models, which considers neighboring satellite fault information for decision making. The diagnostic network does not need to be retrained when the number of formation satellites changes, enabling plug and play. Finally, a federated learning framework is used for distributed training, integrating fault features of the entire satellite group without increasing communication pressure. Each satellite uploads local model parameters for collaborative modeling, improving the fault diagnosis capability for different fault types of satellite group and completing the fault diagnosis. Simulation results demonstrate high accuracy of 99% on the test set, indicating the effectivenessw of proposed method.

The fault detection for quadrotor unmanned aerial vehicle (UAV) is studied in this paper. Considering the UAV model is nonlinear and strongly coupled, a performance driven fault detection method is proposed based on neural network. However, the established fault detection system cannot be applied when the UAV enters a new gravitational field. To deal with solve this problem, a fault detection method is proposed based on transfer learning. By means of subspace transfer method and Bregman divergence measurement method, the source domain and target domain are aligned, and the parameter transfer and threshold setting of neural network are realized. Finally, we verify the effectiveness of the proposed method in a four rotor UAV system.

Aiming at the problem of anomaly monitoring due to the insufficient sensors and low confidence level of the rocket power system, an improved support vector machine based anomaly monitoring method for the rocket power system is proposed. Firstly, a closed loop of the rocket control system is built, and the flight state dataset is constructed by selecting suitable measurable parameters. Secondly, the LSTM Auto encoder algorithm is used to reconstruct the flight state data to obtain the residual data. Then, the support vector machine model is constructed, and the artificial bee colony algorithm is used to find the optimal classification parameters for the support vector machine parameters. The residual dataset is input to the support vector machine model. Finally, the effectiveness and feasibility of the algorithm are verified via the closed loop simulation data.

Anomaly detection of spacecraft telemetry data is a key technology to identify the status of spacecraft and ensure the safe and reliable operation of spacecraft. However, anomaly detection of spacecraft telemetry data usually faces problems such as large dimensionality of time series data, unbalanced anomalies, and lack of labeled samples. In response to these problems, a deep anomaly detection model is proposed based on the idea of anomaly detection. Specifically, according to the strong temporal correlation of telemetry data, a long short term memory network with an attention mechanism is used to establish a telemetry data prediction model. At the same time, in order to overcome the problem of few abnormal labels and high data dimensions of spacecraft telemetry data, a fine tuning transfer learning method is used to optimize the prediction model, and a fully connected layer is used to unify the dimensions of different data sets, by which the accuracy of the transfer learning model and the capacity for anomaly detection are improved. Two spacecraft data sets released by NASA are taken as the experimental object, and the proposed anomaly detection method is used to identify the abnormal state of the data set. The results show that compared with the classic anomaly detection algorithm, the introduction of transfer learning can significantly improve the performance of the model. The experimental results are better than the current common anomaly detection models, which proves the effectiveness of the method.

The remaining useful life (RUL) of the aero engine is important for the safe operation of the engine equipment and the development of maintenance plans. At present, the existing methods are difficult to effectively extract the degradation features of equipment under complex operating conditions and complex faults. To solve this problem, an engine RUL prediction method based on multi scale temporal convolutional network (MTCN) is proposed. In this method, time convolutional networks are used to extract temporal information. Moreover, the degradation features of equipment under complex operating conditions are extracted by multi scale convolution kernel. As a result, it is better to predict the RUL of equipment under extreme conditions. To verify the validity of the proposed method, abundant experiments are carried out on the C MAPSS dataset. The results show that the proposed method can effectively improve the accuracy of RUL prediction under complex conditions.

Health state prediction is a key technology to ensure the safe and stable operation of spacecraft in orbit from a system level. This paper proposes a method for predicting the health status of spacecraft based on unsupervised clustering and long short term memory (LSTM) networks, in response to the characteristic of performance degradation in key components of mechatronics. This method first extracts high dimensional time domain features of multi dimensional parameters of a single component of spacecraft, and fuses them into health factors that reflect the operational status of components through PCA method. Then, it combines unsupervised clustering algorithm to identify different evolution stages of faults. Finally, LSTM network is used to construct a health state evolution prediction model for each degradation stage, achieving health state prediction of spacecraft component. This article takes the key component of the control system, the Control Moment Gyroscope (CMG), as an example to experimentally verify the effectiveness of the above algorithm.

Designing intelligent operating methods is a key for constructing autonomous operating abilities of core devices such as spacecraft. Benefiting from the development of machine learning techniques, current intelligent operating methods driven by data have shown significant improvements on the ability of autonomous operation. However, viewing the trend of spacecraft clusters, traditional methods are challenged by two key requirements, distributed learning and privacy protection. A feasible solution is based on federated learning whose major concerns are how to learn efficiently in a distributed way with privacy performance guarantee. Core devices like spacecraft usually work in extreme environments and are very limited on the resources of computation and communication, and different devices show significant heterogeneous characteristics on data distributions, computation resources and so on. The heterogeneous characteristics can reduce the performance of general federated learning methods. Therefore, in this paper, based on the idea of grouping models, a federated learning algorithm for constructing intelligent operation methods is proposed, which is designed with consideration of the heterogeneous characteristics. The proposed method can reduce the waiting costs among different heterogeneous devices, adjust the timing of local learning of different devices, provide different models for devices with significantly different data distributions, and achieve the goal of improving the performance of operation models obtained by federated learning. Experimental results are conducted to show the the effectiveness of the proposed method

Accurately predicting the remaining useful life (RUL) of lithium batteries plays an important role in understanding their health and managing spare parts resources. Most of the existing lithium battery remaining life prediction methods are limited to the prediction results based on the number of cycles. It is essentially a method oriented to a single time scale, ignoring the practical problem that the health state of lithium batteries is affected by the dual time scales of cycle times and working time. In view of this, this article proposes a lithium battery RUL prediction model based on Transformer under the dual time scales. This method selects the capacity as a key index to characterize its performance degradation. The battery capacity data is processed to obtain training sets and test sets through Kalman filtering and sliding time window. The life information contained in the dual time scales, and fully consider the interrelationship between the life information of different time scale, further, establish a mapping relationship between the capacity and the dual time scales, so as to realize the accurate prediction of the RUL of the lithium battery at the dual time scale. Finally, the effectiveness and potential application value of the proposed method are verified by lithium battery examples.