Yuehua Cheng

Satellite autonomous operation requires fault diagnosis results with evidence paths and structured reports. This paper proposes a knowledge graph-enhanced large language model framework for satellite component fault diagnosis. A data-driven classifier first generates ranked candidate faults, while window-level telemetry features are mapped to evidence keys for graph retrieval. Bidirectional reciprocal ranking fusion integrates classifier ranking, forward graph reasoning, and backward evidence verification to produce traceable diagnostic candidates. Constrained large language models (LLMs) then generate schema-compliant diagnostic reports. Experiments on simulated solid-state power amplifier and reaction-wheel datasets show that the framework preserves data-driven diagnostic performance while improving evidence traceability and structured reporting.

The air data system (ADS) consists of multiple sensors designed to measure atmospheric parameters during an aircraft’s flight. These measurements are then processed by the air data computer to compute the relevant flight parameters required for guidance and control. However, due to the system’s inherent complexity and the harsh flight environment, internal sensors are highly susceptible to failures, including scenarios where multiple components fail simultaneously. To address the issue of complex fault diagnosis in the aircraft air data system, this paper investigates a fault diagnosis method based on the unscented Kalman filter (UKF) and capsule network (CapsNet). First, a UKF model for atmospheric parameters is established based on the computational principles of the air data system, and parameters that can effectively reflect fault information are selected. Then, feature mode decomposition is applied to enhance fault characteristics, which are used as inputs for the neural network. Next, the convolutional layers of the CapsNet are improved using the InceptionV3 network to enhance feature extraction capabilities, and a multi-label classifier is designed to achieve fault location of compound faults. Finally, the proposed algorithm is validated on the simulation platform. Through ablation and comparative experiments, the proposed algorithm achieves a fault diagnosis accuracy of over 95%, demonstrating a more effective solution to the compound fault problem in the aircraft air data system. In our study, compared to convolutional neural networks (CNN), fully convolutional network (FCN), temporal convolutional network (TCN), Inception V3 and the original CapsNet, the improved CapsNet achieved the best performance in compound fault diagnosis.

High-power electric air heaters are key charging components in air thermal energy storage systems, but the dominant heat-loss regions and retrofit basis of existing devices remain unclear. In this study, a three-dimensional conjugate heat-transfer model was developed for an existing 1200 kW vertical electric air heater and validated using three steady-state experimental cases, with a maximum outlet-temperature deviation of 2.17%. Based on the validated model, temperature-field characteristics and segmental heat-loss distributions were analyzed under different mass flow rates. The results show that heat loss was highly non-uniform: Segments 2 and 3 accounted for 37.26% and 54.51% of the total heat loss, respectively, contributing 91.77% in total. A targeted local retrofit scheme was, therefore, proposed by filling the non-flowing inner-cylinder region in Segments 2 and 3 with glass wool and enhancing insulation near local cooling boundaries. After optimization, the average total heat loss decreased from 31.94 kW to 17.69 kW, corresponding to a 44.6% reduction. Under the rated condition, the outlet temperature increased from 1421.6 K to 1482.0 K, providing 584.8 kWh of additional effective thermal storage per cycle and an estimated payback period of 399 d. This study provides a diagnosis-guided retrofit approach for existing high-power electric air heaters.

To enhance actuator fault detection capabilities of unmanned aerial vehicles (UAVs) under complex conditions such as varying wind speeds and strong disturbances, this paper proposes a domain generalization method based on an improved Deep Subdomain Adaptation Network (DSAN). Compared to existing domain generalization methods, the improved DSAN introduces the following advancements for superior cross-domain robustness: firstly, it employs Local Maximum Mean Discrepancy (LMMD) to align subdomain feature distributions between source and target domains, effectively adapting to intra-class distribution variations, outperforming traditional global domain alignment approaches. Secondly, it utilizes Maximum Mean Discrepancy (MMD) to increase inter-class feature distances, enhancing fault classification robustness and reducing false alarm rates. To address unknown target domain distributions, an ensemble domain discriminator with a teacher-student architecture, combined with knowledge distillation, is introduced to assess distribution uncertainty, significantly improving adaptability to unseen domains compared to conventional domain generalization strategies. A UAV flight testing system was developed to validate the method, with multi-scenario experiments conducted under various wind conditions and fault types. The dataset covers normal and faulty flight states in environments including calm, light breeze, moderate wind, and strong wind. Experimental results show consistently higher fault detection rates and lower false alarm rates across tasks, significantly outperforming traditional LSTM and original DSAN. These findings demonstrate strong cross-domain robustness and promising engineering applications.

In multi-agent systems (MASs), the information connections between agents present significant challenges for the fault detection of the system. This paper studies MASs from the perspective of information, constructs and explores the significance and connotation of the information Lagrangian, and relates it to the operational status of MAS as an index for detecting whether a fault has occurred in MAS. In this process, the distribution of the potential function on MAS information manifold is analyzed, the relationship between geodesics, Fisher information distance (FID), and information Lagrangian on the information manifold is studied. The distribution of the MAS information Lagrangian is explained in depth from a perspective of information amount and least action principle. The simulation application verified the effectiveness of the proposed method. As an operating status and fault detection index of MAS, information Lagrangian has clear physical significance, and has scalability potential.

This article proposes a new hierarchical Bayesian semantic-guided attribute transfer (HBSAT)-based data-physics dual-driven fault diagnosis (FD) method for uncrewed aerial vehicle (UAV) swarm systems with unseen faults. First, based on the designed hierarchical fault attributes of UAV swarm systems, the HBSAT is developed to progressively learn highly matched correspondences between fault features and attribute semantics from available fault samples for learning attribute knowledge and attribute-related feature representations, which can be transferred to diagnose unseen faults. Furthermore, a mathematical model of the UAV swarm system is established to generate simulated unseen fault data consistent with fault attributes, which can help the FD model to learn more features and attributes related to unseen faults and improve the diagnostic performance. Besides, the proposed network is extended into the Bayesian deep learning framework to quantify uncertainty. The validity and advantages of the proposed approach are verified based on a semiphysical platform of a fixed-wing UAV swarm system.