Enhancing satellite telemetry reliability is essential for status assessment and operational management. However, telemetry data have complex spatio-temporal dependencies and diverse missing patterns, making it challenging for traditional methods with discrete modeling, fixed parameters, and static prediction steps to recover. To address these, this paper proposes a continuous spatio-temporal information redundancy network that transforms degraded discrete data recovery into timestamp-aware continuous-trajectory inference with uncertainty-calibrated selection. Specifically, a long short-term memory network is reconstructed using neural stochastic differential equations to model hidden-state evolution in continuous time, and Ito integrals are incorporated to characterize uneven observation gaps, thereby enabling recovery at arbitrary timestamps. Additionally, time-varying signals and Brownian perturbations are integrated into the model to allow dynamic adjustment of parameters with the input, thereby improving the ability to recover multiple evolving features. Finally, a structure-prior-guided hierarchical graph is constructed to fuse multi-scale dependencies while suppressing redundant error propagation, and multi-head graph attention is used to derive adaptive confidence intervals for selective recovery, thereby improving telemetry reliability without obscuring novel patterns. Experiments on five datasets show the proposed method improves recovery accuracy by 27.7% over the best baseline, achieves 15.2-78.5 ms batch latency, and has strong potential for real-time large-scale telemetry reliability.

As the heart of an aircraft, the operational condition of an aero-engine ensures flight reliability and passenger safety, so it is critical to monitor its conditions. The prediction-based condition monitoring methods have received more attention due to their advantages of adaptability and high accuracy. In particular, multi-step prediction can monitor aircraft operational conditions in advance to make timely decisions. However, the prediction accuracy decreases as the prediction steps increase due to the lack of extraction capacity and the accumulation of errors; it is significant to improve the precision of the multi-step prediction. Thus, this paper proposes an Informer-based multi-step prediction method for aero-engine condition monitoring. Firstly, data preprocessing is performed, including missing value handling, standardization, correlation analysis, and data reconstruction to obtain high-quality data for accurate model building. After that, an Informer model is built and optimized for multi-step prediction for timely monitoring of flight conditions, which improves the primitive self-attention mechanism and increases model accuracy and inference speed. Finally, the real flight data is applied to detect the superiority of the proposed approach. The experimental results show that the Informer method outperforms the comparative approaches for multi-step prediction of aero-engines, which provides a timely and reliable scheme for condition monitoring.

Phase difference estimation is a core component of measurement and calibration in multichannel systems. For two channels with the same nominal frequency, a joint sinusoidal model with seven parameters is employed to estimate the phase difference, thereby mitigating estimation bias from minor frequency mismatches. The parameters are obtained by solving the least-squares (LS) problem via singular value decomposition (SVD). Analysis shows that when the system matrix has a high condition number (CN), matrix computations are prone to ill-conditioning, leading to numerical instability in the iterative process. To alleviate this problem, this article derives a closed-form scaling factor that minimizes the system matrix CN while preserving the LS solution, thereby ensuring numerical stability. Based on different amplitude relationships between channels, the CN under coherent sampling can be optimized to the range of (14, 110). Simulation results show that the ill-conditioned risk of the original SVD-LS algorithm increases significantly as the CN enters the medium-to-high range. After introducing the scaling factor derived in this article, the CN of the system matrix decreases significantly, the iterative stability and convergence rate are systematically improved, and the LS optimal solution remains unchanged.

Unmanned Aerial Vehicles (UAVs) have been widely used in various industries due to their convenience. However, misuse may pose threats to society. Therefore, anti-UAV detection in low-altitude security scenarios is important. The existing detection methods have a core contradiction of being difficult to balance detection accuracy and detection speed, and the detection effectiveness will be obviously reduced in scenarios with complex backgrounds and extremely small UAV targets. To address these issues, this paper proposes an enhanced anti-UAV detection method DCR-YOLO, which improves the YOLO11 model by introducing a triple collaborative optimization strategy. First, the DySample upsampling module is integrated to retain more effective information in the shallow features and make up for the defect of sparse features of small targets. Second, the Convolutional Block Attention Module (CBAM) is incorporated into the detection head to suppress complex backgrounds and enhance the ability to extract features of small targets. Finally, the lightweight ReGhostConv convolution module with channel distillation effect is introduced, which not only can effectively accelerate detection speed, but also can reduce computational complexity and minimize the model size. Experiments on two open-source detection datasets and tracking datasets show that, compared with mainstream detection methods, the proposed method significantly improves the detection accuracy and speed of small target UAVs. Meanwhile, the parameter has been reduced, achieving lightweight. Detection accuracy and speed have been effectively balanced.

Accurate remaining useful life (RUL) prediction of lithium-ion battery packs is important for reliable operation and maintenance. However, pack-level lifecycle degradation data are costly and time-consuming to obtain, which limits the development of data-driven prediction models. In addition, pack degradation is not directly equivalent to cell degradation because it is affected by inter-cell inconsistency and series constraints. This discrepancy further increases the difficulty of RUL prediction with limited pack-level data. To address this problem, this paper proposes a cross-level RUL prediction method using prior cell degradation and limited pack-level data. Full-lifecycle cell data are first used to establish long-timescale state of health (SOH) trend priors and short-timescale degradation-rate priors through a multi-kernel deep Gaussian process. Limited early-stage pack-level data and a voltage-derived inconsistency feature are then introduced to calibrate the discrepancy between cell-level priors and pack-level degradation behavior through cross-level scaling and residual compensation. During recursive prediction, particle filtering is used to fuse the corrected SOH prediction and degradation-rate prediction, so that the pack-level SOH trajectory and RUL can be obtained. Experimental results on laboratory and electric-vehicle battery pack datasets show that the proposed method achieves an average RUL prediction error of about 4 cycles when only 17.46% of the pack-level lifecycle data are used for model construction. These results indicate that the proposed method can reduce the dependence on costly pack-level lifecycle data while maintaining accurate pack-level RUL prediction under limited-data conditions.

Composite overlapping defects, occurring between the interfaces in multilayer composite, bring great challenges for accurate structural integrity evaluation due to the complexity and unpredictability. Terahertz (THz) technique, as a novel nondestructive testing (NDT) method, has emerged great potentials for complex multilayer composites. However, current THz feature extraction methods, including physics-driven and pure data-driven methods, mainly focus on the identification of simple defect form, such as single non-overlapping defect, while ignoring the applicability of multiple overlapping defects. In addition, the variations of THz testing conditions can cause the variations of THz signals from the same sample, which will inevitably degrade the generalization performance of methods. Therefore, in this work, a physics informed THz multi-label dual adversarial network (PI-THzMDANet) is proposed for intelligent identification of composite overlapping defects under variable testing conditions. It includes a data alignment layer, a multi-label classification layer, and the dual adversarial network. Initially, the data alignment layer is constructed to align the different temporal THz signals in data space based on the transmission and variation characteristics of THz wave. Then, the multi-label classification layer with multi-attention class encoding module (MACE) is innovatively proposed to accommodate the output form of composite overlapping defects. Finally, the dual adversarial network, including the global and local adversarial learning processes, is proposed to improve the generalization performance of PI-THzMDANet under different THz testing conditions by aligning the source and target THz data in latent feature space. The experimental results indicate that the effectiveness of proposed PI-THzMDANet for intelligent identification of composite overlapping defects under variable testing conditions, which provides a novel paradigm for the structural integrity evaluation of complex defect forms in multilayer composites in THz NDT.

Terahertz non-destructive testing (THz NDT) has emerged significant potentials in the thickness measurement of composite structures. Generally, the THz thickness measurement depends on the accuracy of time-of-flight (TOF) extraction from measured THz signals. However, since THz wave is susceptible to the effects of attenuation, dispersion, and multiple reflections, the accurate TOF extraction is often compromised. The present methods include the physics-driven signal processing methods and the data-driven methods. The physics-driven methods rely on the sufficient prior knowledge of THz wave propagation and the manual selection of hyperparameters. Data-driven methods often suffer from the “black box” limitation and lack of physical interpretability. In this work, a physics induced interpretable sparse unrolling network for precise THz thickness measurement of composite structure is proposed. The network aims to solve the sparse inverse problem in THz thickness measurement by unrolling the iterative process of iterative shrinkage-thresholding algorithm (ISTA) into an end-to-end network with physical constraints. The network contains three key modules, including multiple multi-scale convolutional sparse coding (MS-CSC) unrolling layers, effective squeeze-and-excitation (eSE) attention and channel fusion module, and a dynamic sparse-weighted focal loss (DSWF-Loss). MS-CSC unrolling layers are used to perform the iterations, corresponding to the iterative process of ISTA. For each MS-CSC unrolling layer, a physics inspired relaxed dictionary is introduced to improve the interpretability and prediction ability of network. In addition, DSWF-Loss is specially designed to improve the feature capture ability of network for the non-zero elements in sparse TOF. Finally, a series of numerical simulations and comparison experiments are conducted to verify the effectiveness and advantages of the proposed network on three different composite structures. Overall, this work provides a novel insight for the accurate THz thickness measurement of composite structures, which will promote the application of physics-inspired interpretable network in THz NDT.

State-of-health is a critical indicator quantifying the performance degradation of satellite lithium-ion batteries. Due to the strictly limited operating conditions, extracting the health indicators based on the measurable parameters and mapping them to the battery capacity is the only approach estimating the on-orbit state-of-health of satellite lithium-ion batteries. In-orbit data typically lacks true capacity labels, constraining SOH estimation accuracy. Conversely, ground accelerated tests yield labeled data but under stress conditions that differ markedly from on-orbit environments. Bridging the gap between models trained on ground-based multi-condition data and the specific requirements of on-orbit estimation remains a critical challenge. To address these challenges, this paper proposes an on-orbit battery state-of-health estimation framework based on multi-operating conditions uncertainty-aware collaborative learning approach with adaptive weight aggregation. Firstly, considering the on-orbit limited operating conditions, three degradation features are extracted and fused to accurately represent the battery capacity fade. Secondly, a Bayesian Long-Short-Term-Memory network is established to represents the state-of-health degradation processes under multi-operating conditions, thereby quantitatively capturing the inherent uncertainties from different charge and discharge stresses. Finally, a collaborative learning method with adaptive weight aggregation is proposed to enable sustained and distributed collaborative training across diverse operational conditions using ground test data. This uncertainty-oriented collaborative learning paradigm allows new data from different operating conditions to be continuously incorporated and more clients to join the model training process. Therefore, the state-of-health estimation model remains adaptability and robustness against evolving degradation patterns. The performance of the proposed method is validated on both on-orbit simulation data and public datasets. It achieves the mean absolute error of 0.56% under on-orbit operating conditions. These results demonstrate that the proposed method offers a promising and practical solution for satellite lithium-ion battery state-of-health estimation.

With the rapid growth of the global commercial space industry, the satellite power system, as a critical component for ensuring stable satellite operation, is under increasing pressure to meet high standards of reliability and cost efficiency. Extended operation in extreme space environments often gives rise to incipient faults, which evolve gradually and exhibit weak observability, making them difficult to detect. These incipient faults, if not identified and addressed in time, can accumulate and escalate into severe failures, posing risks of mission loss and significant economic impact. However, traditional fault detection methods still struggle to effectively capture such subtle anomalies, highlighting the urgent need for more sensitive and intelligent approaches to incipient fault detection in on-orbit satellite power system. To address this issue, this work proposes a spatiotemporal feature construction and enhancement method based on satellite power system telemetry data to improve the detection capability for incipient faults. This work first captures the long-term dependencies and global correlations in telemetry data to extract effective features of incipient faults. Then, a hybrid multi-scale CNN-Mixer network is designed to further enhance feature representation and eliminate redundant information, thereby improving the accuracy and generalization performance of incipient fault detection. Experiments across 11 operating conditions achieve an accuracy of 1.00 (95% confidence interval: 0.99-1.00) and a false-alarm rate of 0.00% (95% confidence interval: 0.00%-0.47%), demonstrating excellent detection performance and robustness for on-orbit satellite power systems.

Robust prediction is essential for ensuring the reliability of aircraft operations and mitigating safety risks in complex, dynamic environments, as required by condition-based maintenance. However, maintaining generalization across dynamic phase transitions and efficiently quantifying uncertainty remain significant challenges. To bridge this gap, this paper proposes a knowledge-enhanced independent subnetwork (KISNet) model for robust flight data prediction. Based on a multi-input multi-output independent subnetwork architecture, this model efficiently produces predictions with uncertainty estimates in a single forward pass and integrates aviation domain knowledge in two aspects. In data representation, a variable-scale flight phase subdivision method with directional augmentation is proposed to explicitly incorporate flight dynamic characteristics into input features, effectively overcoming cross-phase data distribution discrepancies and training sample imbalance issues. In model optimization, we construct a risk-aware optimization mechanism. By embedding phase-dependent dynamics and risk sensitivity into the learning objective, this mechanism steers the model towards physically consistent and robust updates. Extensive experiments based on real flight data from fixed-wing unmanned aerial vehicles (UAVs) demonstrate that KISNet reduces prediction errors while providing reliable uncertainty bounds and interpretability, showcasing superior robustness and efficiency in high-risk aviation scenarios. While currently validated on UAVs, this model offers a promising baseline for broader applications.

Real-time anomaly detection for flight control systems (FCSs) based on edge intelligence systems is essential for aerial vehicle safety. Extracting and fusing temporal evolution features and cross-dimensional spatial relations FCSs is an effective approach for anomaly detection. But existing models tend to achieve this through complex model structure with abundant parameters, which leads to substantial computational costs for real-time anomaly detection on resource-limited embedded systems. This article proposes a sparse hierarchical spatial-temporal fusion model for real-time anomaly detection of FCSs. First, a spatial feature extraction network is designed based on feature decoupling attention mechanism and weight-decay random forest to alleviate the dimensionality curse during the spatial correlation extraction. Second, a sparse temporal feature extraction network is designed based on multiscale temporal convolution and model-coupled sparsity regularization to reduce redundant parameters while preserving temporal modeling capability. Finally, anomaly detection thresholds are determined by statistically analyzing the prediction residuals to enhance the adaptability of the proposed method under dynamic operating conditions. The proposed method is validated using two real datasets from satellites and uncrewed aerial vehicles. Experimental results show average anomaly detection accuracy of 98.3% with false alarm rate of 5.0%. The model is then implemented on an embedded intelligence system, achieving an average single-step detection time of 1.22 ms. This demonstrates the potential of the proposed method for real-time anomaly detection in FCSs.

With the continued deployment of large scale low earth orbit (LEO) satellite constellations, intelligent operation and maintenance (O&M) systems increasingly rely on telemetry centric time series data as their primary knowledge source. However, constrained jointly by ground station visibility and communication scheduling, telemetry exhibits pronounced asynchrony: high-frequency real-time telemetry (RTT) can only be acquired within short visibility windows, while low frequency delayed telemetry (DT) suffers from substantial latency, making it difficult to obtain a unified and accurate constellation level state estimate. To address this issue, this work proposes an intelligent state estimation method based on asynchronous telemetry fusion. The method first constructs a physics constrained dynamic graph using orbital dynamics and link reachability to characterize the time-varying constellation topology. On this basis, it designs a dual channel feature extraction architecture for RTT and DT and employs an attention mechanism for adaptive fusion, yielding a compact constellation level state vector. Under both real and simulated operating scenarios of a representative Walker constellation, this module is instantiated as a hybrid telemetry dynamic graph neural network (HT-DGNN) and compared against three representative baseline models. The results demonstrate that the proposed method achieves significant advantages in terms of estimation error, temporal consistency, and robustness to long-term evolution and observation gaps, providing smoother and more reliable state trajectories. This indicates good engineering adaptability and facilitates integration into monitoring, anomaly awareness, and mission scheduling workflows in constellation O&M.

Series-connected lithium-ion battery packs are key energy units in electric and storage systems, and accurate state-of-health (SOH) estimation is important for safe operation and capacity management. However, under the coupled effects of operating-condition variation and inconsistency evolution, point-estimation models are prone to mismatch. Interval estimation can describe this uncertainty, but existing methods often produce overly wide intervals or biased interval centers. In addition, single-dimensional health indicators (HIs) are often insufficient to characterize the complex degradation behavior of battery packs. To address these issues, this paper proposes an SOH estimation method for series-connected lithium-ion battery packs that combines multi-dimensional HIs construction with interval optimization. First, a graph neural network with multi-distance adaptive fusion is used to quantify inconsistency at multiple scales, and multi-dimensional HIs are constructed by combining inconsistency-related, incremental-capacity, and voltage-response features. Second, the Bayesian neural network is used to generate the SOH confidence interval, and the multi-kernel Gaussian process regression model is used to establish the reverse mapping from SOH to HIs. Within the estimation interval, candidate SOH values are iteratively optimized by joint statistical consistency, so that interval compression and interval-center correction can be achieved simultaneously. Experimental results on measured battery-pack data show that the proposed method achieves a mean absolute error below 1.57% and reduces interval width by 75.21% compared with the baseline method. These results demonstrate that the proposed method can effectively mitigate model mismatch and improve pack-level SOH estimation performance.

The time-interleaved analog-to-digital converter (TIADC) is a critical component for high-speed and highprecision signal acquisition. However, channel mismatches (offset, gain, and timing skew) significantly degrade performance, necessitating effective adjustment. Existing frequency-domain adjustment methods based on traditional FFT are sensitive to non-periodic input signal sampling, leaving room for improvement in mismatch error estimation accuracy. This paper proposes a TIADC mismatch adjustment method based on the allphase fast Fourier transform (apFFT). By mitigating spectral leakage and enhancing parameter measurement accuracy, the method ensures robust and precise error estimation. It utilizes apFFT to estimate relative mismatches, compensating for gain and timing skew errors while eliminating offset via a mean filter. Compared with existing approaches, it avoids complex filter design and coefficient solving, achieves faster convergence with fewer samples, maintains high estimation accuracy, relaxes input frequency constraints. Moreover, it requires no reference channel and is scalable to arbitrary channel counts. Verified by simulation and implemented on a developed 8-bit 1GS/s TIADC platform, the proposed method improves SFDR/SNDR by 16.9 dB/10.7 dB for single-tone and 16.9 dB/6.7 dB for dual-tone inputs.

Unmanned Aerial Vehicle (UAV) swarms accomplish complex tasks through coordinated formation with their autonomy, timeliness, and intelligence. Phase identification for UAV swarm formation is a critical prerequisite for subsequent formation monitoring and performance assessment. However, accurate formation phase identification faces two main challenges. First, conventional phase definitions are diverse and complex, making identification difficult. Moreover, existing unsupervised learning methods yield inaccurate results due to insufficient extraction of high-dimension spatial and temporal coupling features of collaborative UAVs. To address these challenges, a UAV swarm formation phase identification method based on knowledge-guided spatiotemporal representation is proposed. Firstly, unified formation phases are defined and hierarchical identification subtasks are decomposed by knowledge-based task decomposition to improve phase separability and reduce potential misidentification. Then, formation phases are identified through a spatiotemporal representation model and a density-driven identification optimization mechanism. The model captures inter-vehicle spatial features, parameter couplings and intra-vehicle temporal features, while the optimization mechanism refines clustering results by mitigating high-frequency switching regions during transition phases. Finally, the proposed method is validated utilizing formation data generated by a UAV swarm formation simulation platform built on opensource software. Experimental results show that the proposed method achieves an average identification accuracy of 94.88% and an average F1 score of 0.965. Compared with other methods, the proposed method improves the average identification accuracy by 5.42% to 16.93% and the F1 score by 0.040 to 0.120, outperforming stateof-the-art methods. Consequently, ablation studies and comparative experiments demonstrate the informative spatiotemporal representation of the proposed method with superior performance.

The aircraft auxiliary power unit (APU) is a small turbine engine that provides power and air sources for the aircraft. Its main role is to help start the main engine and provide electric power to the aircraft. The accurate performance assessment (PA) of on-wing APUs can help improve the safety of APUs while reducing unnecessary maintenance costs for airlines. Due to the hostile operating environment and working conditions, the performance parameters are affected greatly. It is difficult to conduct the PA for on-wing APU. In this article, a multiparameter PA approach based on degradation feature enhancement is proposed to fulfill the PA of on-wing APU. First, an adaptive feature extraction variational mode decomposition is proposed to extract the degradation features from the on-wing sensing data and obtain a feature set of the monitored parameters. Then, the extracted degradation features are fused through a long short-term memory (LSTM) network for achieving PA. To evaluate the effectiveness of the proposed method, experiments are conducted based on real on-wing sensing data from airlines. The PA results show that the proposed approach can obtain better PA results.

Abnormal state assessment of unmanned aerial vehicles (UAVs) is crucial for flight safety and mission accomplishment. However, its practical application faces two significant challenges: the degradation of generalization performance under different trajectories and the roughness of 0-1 binary anomaly assessment. To this end, a variable trajectory-oriented refined anomaly assessment method for UAVs is proposed. This method, containing three distinctive modules, breaks through the limitations of traditional anomaly detection. A similarity-guided trajectory identification strategy is first established. This strategy achieves multitrajectory feature decoupling through feature space matching, enabling the construction of a trajectory specificity baseline model library. Additionally, a trajectory-oriented model adaptive optimization strategy is presented. This strategy employs dual-stage collaborative optimization, incorporating the baseline model's preference and parameter dynamic adjustment, to enhance the generalization ability of models. To derive more refined and reliable assessment results, this study innovatively designs a quantitative assessment mechanism for abnormal states. Based on the ranking of abnormal state scores incorporating data deviation and uncertainty, a fine-grained identification of UAV abnormal states is realized. The experimental results using measured UAV flight data from different trajectories demonstrate that the proposed method significantly outperforms existing approaches in variable trajectory scenarios.

In recent years, lithium-ion battery packs are widely used in several fields. State of health (SOH) of lithium-ion battery packs is a key parameter for evaluating the degradation of their performance. For battery packs, in addition to internal cell degradation, cell inconsistency also has a significant impact on the aging process, leading to more complex degradation and greater uncertainty in the SOH estimation. This paper extracts degradation features based on cell inconsistency and performs uncertainty modelling of battery pack SOH. Six single degradation parameters based on cell inconsistency are extracted using parameters such as monitorable voltage and time interval. In order to reduce the information redundancy of single degradation features and improve the features accuracy, the degradation features are fused using the kernel principal components analysis (KPCA) algorithm. The SOH estimation model of the battery pack is established based on long short-term memory (LSTM) network. The model utilizes a parallel training strategy to quantify the uncertainty in the estimation results. The experimental results based on real battery test data show that the correlation between degradation features and battery capacity is between 0.98 and 0.99, and the error of SOH estimation is less than 0.06, and 95% confidence interval is given.

This paper proposes a novel fusion architecture that combines Liquid Neural Networks (LNN) with multi-head attention mechanisms for accurate State of Health (SOH) estimation of lithium-ion battery packs. The method employs learnable liquid time constants (LTC) that enable dynamic adjustment of memory characteristics based on input temporal patterns, while the multi-head attention mechanism identifies critical time steps that contribute most to SOH prediction. Additionally, we introduce an innovative coverage-averaging mapping strategy that transforms overlapping window predictions into smooth, cycle-level SOH estimates, eliminating the boundary discontinuities commonly observed in traditional sliding window approaches. Experimental evaluation on real battery data demonstrates competitive performance with an RMSE of 0.0072 and R2 of 0.9248. The results demonstrate that the proposed method successfully establishes a mapping relationship between individual cells and the battery pack, and accurately estimates the SOH of the pack.

This paper presents CNN-BisLSTM-Attention, a hybrid model for predicting key parameters of aero-engines. The architecture integrates two core modules: (1) A spatio-temporal feature extraction module combining convolutional neural networks for local spatial patterns and bidirectional stabilized long short-term memory (BisLSTM) for long-term dependencies; (2) A self-attention feature enhancement module introducing a multi-head self-attention mechanism to optimize discriminative feature weights. Real data from aero-engines are applied to validate the performance of the model. Experimental results manifest that the CNN-BisLSTM-Attention model outperforms the comparison models in prediction accuracy. Specifically, for MAE and RMSE indicators, the proposed model improved by 18.6%-46.0% and 14.4%-43.5%, respectively. Additionally, ablation studies prove the contributions of each component of the proposed hybrid prediction model.