Current Issue

2026 Vol. 52, No. 11

Display Method:
Volume 11 Issue E-journal
Volume 52 Issue112026
iconDownload (52801) 203 iconPreview
Review of ADS-B anomaly detection research
YANG Hongyu, ZHAO Ziwen, ZHU Youwen, HU Huiling, HU Juncheng
2026, 52(11): 3677-3698. doi: 10.13700/j.bh.1001-5965.2025.0600
Abstract:

The automatic dependent surveillance broadcast (ADS-B) system, as an important component of aviation surveillance technology, has been widely deployed and applied. The application of the ADS-B system significantly improves the coverage and positioning accuracy of air traffic surveillance. Nonetheless, owing to its open broadcast communication framework, this system is susceptible to numerous security risks including data manipulation, harmful message insertion, and interference assaults, which significantly compromise air traffic safety and control determinations. Anomaly detection for the ADS-B system has gradually become a research hotspot. This paper systematically summarizes and analyzes the current research status of anomaly detection in ADS-B systems at home and abroad. Firstly, it analyzes the causes of abnormal data in ADS-B systems and summarizes the types of attacks that this system is vulnerable to and their impacts. The present state of research on ADS-B anomaly detection is thoroughly examined from two perspectives: the consistency check-based approach and the deep learning-based approach. The merits and demerits of several techniques regarding detection precision, generalisation, and practical implementation are examined, with an analysis and summary of existing ADS-B anomaly detection products in relation to their respective application contexts. Ultimately, a comprehensive synthesis of the current research findings reveals the primary obstacles encountered by contemporary anomaly detection techniques for ADS-B, alongside potential future research avenues and developmental trajectories.

Improved lightweight and efficient EHH-YOLOv8s algorithm for strip steel surface defect detection
LIANG Liming, LONG Pengwei, LU Baohe, OU Yanyan, ZENG Lu
2026, 52(11): 3699-3711. doi: 10.13700/j.bh.1001-5965.2024.0426
Abstract:

To address the issues of low efficiency, and poor accuracy in current steel surface defect detection, and the challenges of large model parameters, low detection accuracy, and poor real-time performance of existing defect detection models, a lightweight steel defect detection algorithm based on YOLOv8s (EHH-YOLOv8s) is proposed. The reconstruction feature interaction module H-C2f, which successfully preserves spatial and channel features, suppresses redundant information, and improves detection accuracy and speed; the Haar wavelet downsampling (HWD) module, which expands the receptive field, reduces model complexity, and speeds up training and inference; and the lightweight EfficientViT network, which serves as the backbone network to reduce model complexity and better handle cross-scale feature information. Experimental validation on the NEU-DET and Severstal steel defect datasets shows that compared to the YOLOv8s algorithm, the EHH-YOLOv8 algorithm achieves an improvement in mean average precision (mAP) of 6.9% and 5.4%, respectively, with reductions in parameters and computations of 4.1×106 and 13.2×109, and achieves detection speed of 136 frames/s and 118 frames/s. These findings show that the algorithm strikes a superior balance between speed, lightweight design, and detection accuracy, offering a dependable reference for high precision and real-time performance on edge devices. Finally, validation on the PCB board defect dataset shows that the algorithm possesses good generalization performance.

Experimental research on airframe noise of large-scale full-scaled civil aircraft model in wind tunnel
ZHOU Guocheng, SUN Yifeng, BAO Anyu, DING Cunwei
2026, 52(11): 3712-3723. doi: 10.13700/j.bh.1001-5965.2024.0660
Abstract:

Carrying out aero-acoustic experiments using a large scale model is an important means to accurately grasp the noise characteristics of civil aircraft airframes and validate the effectiveness of noise reduction strategies. The airframe noise experiment for a large-scale civil aircraft model is conducted in the 8 m × 6 m aero-acoustic wind tunnel using its associated measuring and support equipment. The ground linear array, 1/4 arc array, and horizontal microphone phase array are adopted to measure flyover noise, sideline noise and noise source distribution. The experimental equipment and methods adopted are introduced in this paper, and the influence of wind speed, angle of attack, and configuration on airframe noise is analyzed. According to the test results, civil aircraft airframe noise poses broadband characteristics; different Mach number scaling laws apply for different frequency ranges. The lift device and landing gear both contribute to the overall airframe noise. The airframe noise increases with angle of attack and flap deflection angle, and presents certain directionality in flyover and lateral directions. Source distribution of different configurations has significant differences, and the flap side edge noise becomes major noise source in the high frequency range at large flap deflection angles.

Sentiment representation calibration-based model for image-text sentiment analysis
ZHANG Shunxiang, LIU Jiajia, JIAO Yixuan, ZHANG Yulei, LI Ziliang
2026, 52(11): 3724-3734. doi: 10.13700/j.bh.1001-5965.2024.0545
Abstract:

Exploring the association of multi-modal data for sentiment analysis is made more difficult by the growing trend of social media users expressing their opinions through both text and visuals. Existing research methods lack sufficient interaction between modalities and cross-modal interactions often introduce noise, which leads to sentiment representation bias of learned image and text representations. To address this issue, we introduce a sentiment representation calibration-based model for image-text sentiment analysis. Firstly, we encode the image and text through a feature extraction module to capture image and text features. In order to fully utilize the inter-modal complementary information, the cross-modal attention is then employed in the cross-modal interaction module to carry out bidirectional interaction in the "image-to-text" and "text-to-image" directions. Next, noise is filtered out from the interaction-enhanced features through a gated calibration module to correct the sentiment representation bias in the interaction-enhanced features. Finally, in the sentiment detection module, feature fusion is achieved by using multi-head attention to integrate the calibrated feature representations for sentiment detection. Experimental results on two public benchmark datasets demonstrate that the proposed model outperforms state-of-the-art image-text sentiment analysis methods, such as CLMLF and MVCN, thereby improving the accuracy of sentiment recognition.

Aircraft skin damage LHN-YOLO small target detection algorithm incorporating attention mechanisms
LI Yaohua, CHEN Jie
2026, 52(11): 3735-3744. doi: 10.13700/j.bh.1001-5965.2024.0713
Abstract:

Aiming at the problems of low detection accuracy and a large number of parameters for small targets in aircraft skin damage detection, an LHN-YOLO algorithm incorporating the attention mechanism is proposed based on the YOLOv8n algorithm with aircraft skin rivet damage as the research object. To augment the model’s capacity for detecting small targets, an optimized efficient multi-scale attention (EMA) mechanism is integrated into the backbone, with the EMA mechanism refined through the lightweight convolution LightConv, resulting in the Light-EMA module that enhances detection precision while minimizing computational complexity; concurrently, the Haar wavelet downsampling module and normalized Gaussian Wasserstein are incorporated to supplant the original model’s downsampling module. Distance is introduced to replace the downsampling operation and CIoU loss function of the original model to improve the prediction performance. The results of the experiment reveal that the LHN-YOLO algorithm’s average accuracy, recall, and precision are 92.7%, 87.0%, and 91.7%, respectively, which are improvements of 2.2%, 3.5%, and 2.8% over the original model, alongside a 10% reduction in parameters, significantly boosting the detection of small target damage on aircraft surfaces.

Multi-flight trajectory prediction method in terminal area based on an improved spatio-temporal model
HE Huaiqing, SHI Xiaoshuang, LIU Haohan, HUI Kanghua
2026, 52(11): 3745-3757. doi: 10.13700/j.bh.1001-5965.2024.0540
Abstract:

4D trajectory prediction in this setting is extremely unpredictable and difficult due to the large amount of past trajectory data in the terminal area, its rich spatio-temporal properties, and the sudden changes in trajectory point placements. To address these challenges, we propose a solution based on an improved spatio-temporal model comprising three components. First, we introduce temporal convolutional networks (TCN) and bidirectional gated recurrent units (BiGRU) into the field of flight trajectory prediction, improving the TCN architecture. This leads to the development of a double-layer multi-scale temporal convolution network MMTCN that effectively extracts temporal features from both global and local perspectives. Second, we integrate layer normalization and residual connections into the graph attention network GAT to capture the intricate spatial dependencies among flights. Third, to mitigate data uncertainty stemming from abrupt changes in trajectory point locations, we input the extracted spatio-temporal characteristics into a conditional variational autoencoder (CVAE) to learn the latent representations of flight trajectory data, ultimately generating multiple predictive trajectories. Throughout the training process, a learning rate adjustment strategy is employed to enhance training efficiency. The experimental results demonstrate that the improved model achieves optimal performance on two metrics, with an average displacement error (ADE) of 0.0826 and a final displacement error (FDE) of 0.1111, significantly outperforming other baseline models. Additionally, the suggested model lowers the maximum error on the general aviation dataset by roughly 7.75% for FDE and 9.59% for ADE. This model not only effectively extracts spatio-temporal characteristics but also exhibits training stability.

Ultrasonic transmission testing of bonding interface curing process in Al-CFRP composite structures
BU Fanqiang, LIU Yong, WANG Xingguo, LI Xiaogao, SHEN Guolang, MA Chengwen
2026, 52(11): 3758-3769. doi: 10.13700/j.bh.1001-5965.2025.0007
Abstract:

An ultrasonic transmission testing method was created in response to the difficulties in assessing the bonding states of isotropic-anisotropic composite structures. Based on the transfer matrix method, a dimensionless parameter ωτ was introduced to describe bonding interface states. A theoretical model of ultrasonic propagation was established for the composite structure obtained by resin bonding Al and carbon fiber reinforced polymer (CFRP) to analyze the influence of the bonding state on the ultrasonic transmission characteristics. The results of theoretical analyses show that the transmission coefficient under the same incident angle gradually increases when ωτ increases, and the transmission coefficient peak frequency shifts to higher frequency. Theoretical analyses reveal that the transmission coefficient at a fixed incident angle gradually increases as ωτ increases, and the peak frequency shifts to a higher frequency. An ultrasonic transmission time-domain simulation model for the Al-CFRP composite structure was developed, and the simulation results show that the fast Fourier transform (FFT) amplitude gradually increases with ωτ increasing, and the peak frequency shifts to a higher frequency. The transmission coefficient increases rapidly with ωτ at first, then it increases slowly and eventually stabilizes. The trends in FFT amplitude and transmission coefficient with curing time are compatible with the simulation results, according to ultrasonic transmission testing measurements conducted on Al-CFRP bonded specimens under various bonding interfaces.

Structural optimization of flexible blades for vertical axis wind turbines
WANG Xiaozhe, WANG Yubo, WU Kun, WAN Zhiqiang
2026, 52(11): 3770-3778. doi: 10.13700/j.bh.1001-5965.2024.0444
Abstract:

A vertical-axis wind turbine has a rotation axis that is perpendicular to the ground or the direction of the airflow, allowing it to generate stable power without having to orient it as the wind direction changes. Applying flexible blades can further improve power generation, but there is currently a lack of relevant theories and methods. This article proposes a fluid-structure coupling analysis and optimization design method for the flexible blade based on two-dimensional aerodynamics and simply supported beam theory. First, an effective static analysis approach for simply supported beams under distributed loads is proposed, and flexible blade design and structural simplification are carried out based on the fundamental data of current blades. Then, the two-dimensional viscous/inviscid aerodynamic forces and structural deformation are fitted through curve interpolation to establish a fluid-structure coupling response analysis framework for flexible blades. Finally, the influence of the stiffness characteristics of flexible blades on the periodic power generation of generators is studied. The genetic algorithm and Kriging method are employed to maximize power generation by searching for the optimum stiffness of the flexible blades. The results indicate that the flexible blade structure helps to improve power generation, and this study can provide a reference for the research, design, and further application of vertical axis wind turbines.

Low-light image enhancement algorithm for UAV aerial photography base on low-light instances
WANG Dianwei, ZHAO Wulin, FANG Jie, LI Yuanqing, XU Zhijie
2026, 52(11): 3779-3789. doi: 10.13700/j.bh.1001-5965.2024.0531
Abstract:

A two-stage low-light image enhancement algorithm for unmanned aerial vehicle (UAV) aerial photography is proposed to address low-light image issues, including low brightness, fuzzy details, noise interference, and challenges in obtaining paired low/normal light training data. Low-light image enhancement is a crucial preprocessing component of many computer vision applications. Based on the Retinex theory that images with the same content share consistent reflection components, two low-light images with the same content are used for training to learn a priori information, which does not depend on paired low/normal light data pairs. The first stage features a detail enhancement module (DEM) that captures multi-scale feature information to enhance details and edges in the reflection component. To prevent overexposure in high-contrast areas while improving illumination, the second stage incorporates an illumination enhancement module (IEM) and an illumination attention map. Additionally, a denoise module (DM) is incorporated to suppress noise in the reflection component. Experimental results indicate that our method outperforms current state-of-the-art algorithms in visual quality and objective evaluation metrics across multiple datasets. Moreover, the enhanced images significantly improve the performance of target detection algorithms in low-light conditions.

Improved RT-DETR model for aircraft door detection
ZHANG Changyong, ZHANG Xuancheng, QIAO Yahai, LI Zheng
2026, 52(11): 3790-3800. doi: 10.13700/j.bh.1001-5965.2024.0705
Abstract:

Aircraft door detection is crucial for enabling the automatic docking of airport special vehicles. To address the issues of low detection accuracy and difficult model deployment caused by changes in weather and lighting conditions, interference from body coating, and equipment limitations, an improved real-time detection Transformer RT-DETR based aircraft door detection model is proposed, using a lightweight residual neural network ResNet18 as the backbone feature extraction benchmark network. By introducing the context guided block, the BasicBlock in the backbone network is optimized to enhance the utilization of contextual information and reduce the consumption of computing resources. The AIFI-HiLo module is built in the neck encoder network by incorporating the high-low frequency (HiLo) attention into an intra-scale feature interaction (AIFI) module. This innovation strengthens feature representation, effectively reducing missed and false detections. Replace cross-scale feature fusion module (CCFM) with a two-stage improved high-level screening feature pyramid network (HS-FPN) to enhance the model’s ability to fuse multiscale features. The experimental results showed that the average accuracy of the improved model reached 94.9%, which was 1.8% higher than the benchmark model. The parameter count, floating point operation count, and memory usage were reduced by 41.2%, 40% and 41.5%, respectively. The improved model has the characteristics of small size and high accuracy, making it more suitable for airport special vehicles to carry equipment.

Classification method of rotor drones and flying birds under low signal-to-noise ratio for radar
WANG Xiaoliang, WANG Congsheng, SHI Yuxiang, HE Weikun
2026, 52(11): 3801-3813. doi: 10.13700/j.bh.1001-5965.2024.0585
Abstract:

Aiming at the problem of how to classify multirotor drones and flying birds under low signal-to-noise ratio conditions for low-altitude surveillance radar, an effective target classification method is proposed. The subband statistical features, which are obtained respectively from wavelet packet decomposition (WPD) and time-frequency segmentation (TFS) of the radar echo signals, are extracted first in the method. The feature vectors are then created using the subband statistical features, and the targets are classified using a support vector machine (SVM). The subband energy ratio of wavelet packet decomposition and the subband mean value of time-frequency spectrogram are chosen as the classification features to build the feature vectors that are used by the final support vector machine classifier by examining the variations in classification performance of various statistical features. The experimental results show that the classification accuracy of the proposed method for classifying multirotor drones and flying birds under low SNR conditions is significantly improved compared with other common methods.

Three-dimensional path planning of UAVs based on multi-strategy golf optimization algorithm
SUN Bo, ZHANG Wenpeng, WU Zexuan, SU Yebo, WEI Ming
2026, 52(11): 3814-3823. doi: 10.13700/j.bh.1001-5965.2024.0429
Abstract:

This paper establishes a mathematical model for the three-dimensional path planning problem of a UAV to determine the unmanned aerial vehicle(UAV) trajectory and the actual departure time to minimize the total cost of path length, flight height, threat, and smoothing. This model also takes into account a number of limitations, including obstructions, radio interference, dynamic limited areas, UAV performance, and other elements. A multi-strategy golf optimization algorithm (MSGOA) is designed to solve this problem. Sine mapping is used to generate more diverse initial populations. In the exploration phase of the algorithm, Piecewise chaotic mapping and golden sine strategy are integrated to expand the global exploration capability of the algorithm. In the development stage of the algorithm, the Cauchy variation strategy is used to increase the local optimization ability of the algorithm. The improved algorithm is compared with other five mainstream swarm intelligence algorithms on 10 standard test functions to prove the effectiveness of MSGOA. In order to demonstrate the superiority of the model and algorithm, the impact of various factors on the optimal 3D flight route of UAVs is examined, and the performance differences of several algorithms on this topic are compared, using the digital elevation model (DEM) of Hong Kong, China as an example.

Three-dimensional track planning for multi-UAV approach in free airspace environment
WANG Jianzhong, LIANG Feida
2026, 52(11): 3824-3831. doi: 10.13700/j.bh.1001-5965.2024.0496
Abstract:

In order to prevent congestion during regional approach phases, this study suggests a conflict-free multi-unmanned aerial vehicle (UAV) track planning strategy for metropolitan spatial environments under free airspace conditions. The method aims to ensure safe and efficient UAV approach operations while enhancing airspace utilization efficiency. Firstly, we defined the spatial scope of the approach phase. Secondly, we discretized the three-dimensional operational environment using the grid method and introduced the concept of grid occupancy time. UAV acted as dynamic obstacles, and we developed a conflict-free track planning model to minimize flight time. Subsequently, we enhanced the A* algorithm based on a three-dimensional congestion weight map to compute congestion-avoiding flight tracks for multiple UAV. Finally, we simulated approach scenarios with 10, 15, and 20 UAV under various obstacle environments. With cost weight coefficient values of 0.4, 0.6, the results demonstrate that in a complex obstacle environment, the improved algorithm decreases congested areas by 52% while sacrificing just 0.96% of flight time when compared to the A* algorithm. This significant optimization underscores the effectiveness of the model, confirming both its feasibility and the efficacy of the algorithm.

Aerial image small object detection algorithm based on state space model
TENG Shangzhi, GONG Xi, LI Baoan, LYU Xueqiang
2026, 52(11): 3832-3841. doi: 10.13700/j.bh.1001-5965.2024.0561
Abstract:

Small object detection aims to solve the difficulties of traditional target detection algorithms in dealing with tiny objects. Small object detection needs to overcome the problems of small object size, few pixels, and a small difference between the target and background, which makes it difficult to detect. It puts higher requirements on the robustness and accuracy of object detection algorithms. A small object detection algorithm based on a state space model is designed to explore the application of the state space model in the field of object detection. A sliding window scanning mechanism is designed to convert two-dimensional images into one-dimensional sequences, which is more suitable for small object detection tasks. In order to suppress irrelevant background information, a dual-branch structure is used in a spatial information aggregation module to calculate global and local attention, respectively. Improvements are made to the traditional IoU loss, focusing on the influence of bounding box shape and scale on detection performance, and auxiliary boxes are used to accelerate model convergence. Sufficient experiments are carried out on the public dataset VisDrone2019 to demonstrate the effectiveness of the proposed method.

Swin Transformer-based semantic image transmission with model division multiplexing
PENG Yi, SUN Chang, YANG Qingqing, LI Hui, WANG Jianming
2026, 52(11): 3842-3851. doi: 10.13700/j.bh.1001-5965.2024.0542
Abstract:

Communication efficiency is greatly increased via semantic communication, a new paradigm for next-generation communication systems. However, previous works on semantic image transmission were mostly built upon convolutional neural networks (CNN). Traditional CNN-based models often struggle to comprehensively extract the semantic information embedded in images. To address this limitation, this paper adopts the Swin Transformer as the backbone network to effectively capture semantic information. Based on this, a novel semantic image transmission method leveraging model division is proposed. This method achieves the independent transmission of public and personalized features through feature separation. Specifically, after the input image undergoes joint source-channel coding (JSCC), the encoded semantic features are categorized into public features and personalized features, which are then transmitted separately. By enabling resource sharing for public features, the proposed method significantly reduces bandwidth requirements, thereby improving transmission efficiency. To evaluate the performance of the proposed model, experiments are conducted on the publicly available CIFAR10 dataset, comparing the proposed method with traditional approaches. The experimental results demonstrate that the proposed method effectively improves image transmission quality. Furthermore, it achieves superior performance in terms of transmission efficiency and bandwidth savings. In order to overcome bandwidth limitations in semantic picture transfer while preserving high-quality image reconstruction, this work offers a promising alternative.

Refined prediction of terminal area traffic congestion based on multimodal spatio-temporal feature fusion
LI Shanmei, SONG Sini, WANG Hongyong, WANG Duanyang, ZHAO Mo
2026, 52(11): 3852-3863. doi: 10.13700/j.bh.1001-5965.2024.0557
Abstract:

In order to enhance the accuracy and refinement of grasping the future traffic dynamics in the terminal area, a set of traffic congestion prediction methods based on the fusion of multimodal spatio-temporal features is proposed. First, from the standpoint of aircraft detention and control load, a traffic flow clustering technique is used to provide a traffic congestion index calculation method that takes traffic flow heterogeneity into account. Further, the airspace of the terminal area is divided into four-dimensional spatio-temporal blocks, and the spatio-temporal blocks are used as the nodes for constructing the dynamic adjacency network and the traffic flow network, which characterize the spatial correlation of the terminal area in a multimodal way. Finally, a three-dimensional convolutional neural network (3DCNN)+graph convolutional network (GCN)+gated recurrent unit (GRU)-based traffic congestion prediction model for terminal areas is proposed. The model validation was performed using Beijing terminal area trajectory data, and the results show that the method of this paper is superior to the classical baseline model, with the best prediction effect at 30 min compared to 45 min and 60 min time scales, and the root mean squared error (RMSE) and mean absolute error (MAE) compared to the history average (HA) model, exponential moving average (EMA) model, and support vector regression (SVR) regression model and MAE decreased by 21% and 57%, 38% and 72%, and 19% and 51%, respectively. The RMSE and MAE of this paper’s model fall by 36% and 40%, 6% and 8%, and 2% and 2%, respectively, as compared to learning only neighbor space features, learning only traffic flow space features, and transitioning to long short-term memory (LSTM) to learn temporal characteristics. At the same time, the prediction results show that the more complex the trajectory is, the greater its prediction error is, and with the rise in the altitude of the airspace, the prediction error shows an increase and then a decrease. Meanwhile, the prediction results show that the more complicated the trajectory is, the larger the prediction error is, and with the increase of airspace altitude, the prediction error shows an increase and then decrease.

Graph neural network recommendation algorithm fused with dual-channel attention and time encoding
BAI He, LI Yu, HAO Ming, XU Tingting
2026, 52(11): 3864-3878. doi: 10.13700/j.bh.1001-5965.2024.0795
Abstract:

In response to the problem that most graph neural network recommendation algorithms do not incorporate temporal information, leading to wasted information and inaccurate recommendations, this paper proposes a graph neural network recommendation algorithm, GUITRec, that fuses dual-channel attention with time encoding. Utilizing all of the data in the dataset, this model creates a graph structure based on the user’s attributes, the item’s details, and the user-item interaction behavior. Data aggregation is carried out using graph neural networks to create user and item representation behaviors, more thoroughly capturing the traits and connections of both. Secondly, the dual-channel attention mechanism convolutional block attention module (CBAM) is introduced to assign different weights to the user and item representation behaviors, enabling the model to learn according to those weights and pay more attention to important features. Finally, a time encoder model is introduced to encode the interaction time between the user and the item and the user’s location, and new daily, weekly, and monthly cycle encodings and location encodings are added to fully mine time and location information and provide a richer context for recommendations. Experiments on the Ciao dataset show that GUITRec is superior to the baseline model in terms of the three evaluation indicators of mean absolute error (MAE), mean squared error (MSE), and root mean square error (RMSE).

Interactive design and implementation of cluster-oriented industrial CFD software
ZHANG Fan, DING Mingsong, CHEN Jianqiang, LIU Wan, JIANG Tao, LI Peng, JIANG Jun
2026, 52(11): 3879-3893. doi: 10.13700/j.bh.1001-5965.2024.0810
Abstract:

In the field of aerospace engineering, industrial computational fluid dynamics (CFD) software relies on high-performance computing clusters to generate aerodynamic data, serving as a critical tool for assisting aircraft design. This paper devises and executes a CFD software framework for cluster interaction to address the challenges faced by domestic CFD software in high-performance computing cluster deployment and simulation, along with compatibility issues with software/hardware environments, while considering the requirements of the National Numerical Windtunnel (NNW) project for software research, development, and promotion. The architecture employs a client/server model, incorporating features such as an interactive interface, cluster communication, login authentication, simulation task management, and job monitoring, while facilitating a command-based solver invocation and parameter file transfer, demonstrating significant versatility and scalability. Testing outcomes indicate that the framework is congruent with prevalent cluster architectures and job scheduling systems; in contrast to conventional coupled architectures, it diminishes memory utilisation by 36% and CPU consumption by 69% under hundreds of simultaneous tasks, while maintaining average response times within 100 milliseconds, highlighting its benefits in resource efficiency and swift responsiveness. Currently, the designed framework has been applied to the NNW software ecosystem, provide users with smooth cluster simulation interaction services and significantly improve the user experience of CFD simulations.

GTA-Seq2Seq multi-step 4D trajectory prediction model based on spatio-temporal feature extraction
SHI Qingyan, ZHAO Wantong, HAN Ping
2026, 52(11): 3894-3906. doi: 10.13700/j.bh.1001-5965.2024.0679
Abstract:

Real-time flight path correction, conflict avoidance, and increased flight safety are all made possible by high-precision multi-step trajectory prediction. Traditional models usually face problems such as low multi-step prediction accuracy and difficulty in capturing the nonlinear relationship between spatial and temporal features in flight trajectories. Therefore, this paper proposes a multi-step 4D trajectory prediction model GTA-Seq2Seq that uses a gate recurrent unit (GRU), temporal convolutional network (TCN) and temporal pattern attention (TPA) mechanism in a sequence-to-sequence (Seq2Seq) model. The model is divided into an encoder part and a decoder part. The encoder extracts the time dependence in trajectory data by using GRU and combines the TCN-GRU network to extract the spatial features of trajectory data. To accomplish high-precision multi-step trajectory prediction, the spatial and temporal information are concatenated in the decoder section and sent into GRU. The experiment shows that the proposed model outperforms traditional models in terms of prediction accuracy. The fifth step prediction in the model reduces the root mean square error (RMSE) by 27.70% and the mean absolute error (MAE) by 23.82% compared to the optimal model—spatio-temporal encoder-decoder (STED) among the comparison models. In addition, the required number of parameters of the GTA-Seq2Seq model in the training process is reduced by 67.42% compared to STED model, improving the operational efficiency of the model.

Steady-state enhancement design of IMM algorithm for space maneuvering target tracking
ZHANG Shiyuan, LU Shan, LI Qing
2026, 52(11): 3907-3917. doi: 10.13700/j.bh.1001-5965.2024.0746
Abstract:

The relative navigation system for monitoring non-cooperative maneuvering targets in space seeks to acquire precise real-time orbital data and enhance spacecraft’s rapid situational awareness and accurate decision-making for target orbital maneuvers, forming the foundation of the spacecraft pursuit-evasion strategy. The research examines the interacting multiple model (IMM) algorithm for assessing the relative motion state of an unidentified target throughout a sequence of maneuvers, and establishes a model set that comprises a non-cooperative model, a uniform acceleration maneuvering model, and a big deviation response model. Based on the designed model set, the estimation performance factors of IMM are analyzed when the target is in a stable motion of a specific maneuvering mode (steady-state mode). The expression for the equilibrium point of matching model weights (the steady-state weights) in a steady-state mode is established, and the pattern of change of steady-state weights around the equilibrium point is presented. Besides, the influence mechanism of model transition probability and model likelihood ratio on the size of steady-state weights and the sensitivity of measurement noise is clarified. The above analysis provides a theoretical basis for the steady-state performance analysis and optimization of IMM. Based on the above analysis, the correction of the model transition probability is taken in real-time, combined with the development trend of model weights in a given time window. Then, the enhanced steady-state IMM (ES-IMM) is proposed. It not only boosts the weight of the matching model in steady-state estimate but also mitigates weight fluctuations caused by measurement noise, hence improving the steady-state estimation efficacy of IMM in tracking space maneuvering targets. The simulation results show that the estimation accuracy of ES-IMM is better than that of standard IMM in the steady-state interval of space maneuvering target tracking task.

Semantic segmentation algorithms of aerial images via context-aware attention guidance
LIU Lulu, XIE Ronglei, JI Jiachi
2026, 52(11): 3918-3928. doi: 10.13700/j.bh.1001-5965.2025.0356
Abstract:

Semantic segmentation algorithms for aerial images have significant research and application value in fields such as land use classification, environmental monitoring, resource surveying, and urban planning. To address the challenges posed by high-resolution aerial images—such as complex spatial relationships among objects, varying target scales, large intra-class variations, and high inter-class similarities—this paper proposes a context attention-guided semantic segmentation algorithm. In order to improve the network's ability to represent features, a collaborative attention module is added to the encoder stage of the model. This module pulls features from both spatial and channel dimensions. To efficiently extract both semantic information and detailed features, a global-local feature fusion module is built in the decoder stage using concurrent global and local branches. Furthermore, a crisscross semantic interaction module is employed to fuse multi-level features, improving the model’s ability to distinguish object boundaries. Experimental results demonstrate that the proposed algorithm achieves a mean intersection-over-union(MIoU)of 71.0% on the UAVID dataset, significantly outperforming state-of-the-art algorithms in both accuracy and inference speed.

A robust guidance method for reusable rocket landing phase based on UMPSP
ZHANG Yonghua, WANG Zheng, WANG Shiyu, BAI Yunfei, QIU Likuan
2026, 52(11): 3929-3940. doi: 10.13700/j.bh.1001-5965.2024.0491
Abstract:

This paper presents a robust guidance approach for the landing phase, based on unscented model predictive static programming (UMPSP). It seeks to solve the problem of reusable rockets making accurate vertical soft landings in the face of unpredictable atmospheric wind fields. Firstly, the dynamic equation of rocket thrust and attitude coupling is established. Under the consideration of an unknown uncertain wind field, a set of control strategies insensitive to the uncertain wind field parameters of the original system is derived using unscented optimal control and model predictive static programming (MPSP) theories. This aims to minimize the terminal state error dispersion and enhance landing precision. Secondly, capitalizing on the features of the UMPSP algorithm, a rapid analytical guidance methodology is proposed to initialize the control strategy, thereby further boosting solution efficiency. The results of the simulation show that the suggested UMPSP-based robust guidance approach for the landing phase exhibits excellent robustness and superior computing efficiency against a variety of uncertain wind field conditions, while guaranteeing high-precision landing of reusable rockets.

UAV online path planning based on POMDP and parameterized MPC in uncertain environments
YUAN Jinbiao, LIU Haiyan, MA Bodi, ZHAO Leyang, GAO Honggang, LIU Zhenbao
2026, 52(11): 3941-3957. doi: 10.13700/j.bh.1001-5965.2024.0662
Abstract:

It is challenging to eliminate positioning and perception ambiguities when unmanned aerial vehicles (UAVs) operate autonomously in unfamiliar areas. In view of this, it is necessary to conduct research on path planning to enhance the effectiveness of the path under the interference of uncertain factors and thereby improve the target tracking ability of UAVs. The goal of applying model predictive control(MPC)to the real-time path planning system is currently accomplished by Tube-based-MPC, which supports the separation architecture of planning and control, and nominal belief state optimization (NBO), which offers a theoretical foundation for the uncertainty of target tracking. On this premise, this paper proposes an online path planning framework to solve the problems of target tracking and obstacle avoidance in unknown environments. First, the path planning problem is abstracted into a partially observable Markov decision process (POMDP) model, and the Kalman filter is used to estimate the belief state, integrating the uncertainty measure of target positioning into the framework. Secondly, based on the simplified 3-dimensional motion model, a tightened constraint Tube-based-MPC optimization framework is constructed to solve the POMDP. The path planning for a specific continuous time span in the future is achieved by building a nonlinear optimization problem that includes the belief state and unknown barriers. Then, the shapes of UAVs and obstacles are fully considered in the optimization problem, and a full-dimensional obstacle avoidance distance metric is introduced to significantly expand the feasible region of the optimization problem. Finally, the Q-learning reinforcement learning framework is introduced, and MPC is used as the policy function approximator to improve the closed-loop performance of target tracking. Simulation experiments and hardware-in-the-loop experiments verified the effectiveness of the proposed method.

Risk-constrained safe path planning for unmanned aerial vehicles in urban airspace
ZHU Yuanjun, LI Yan, ZHANG Xuejun, ZHANG Weidong
2026, 52(11): 3958-3968. doi: 10.13700/j.bh.1001-5965.2024.0843
Abstract:

This study addressed the safe path planning problem for unmanned aerial vehicles (UAVs) in urban airspace to support the development of the low-altitude economy. A grid-based modeling approach is used to represent the discrete three-dimensional space. A three-dimensional safe path planning model is created to reduce the total path cost, taking into account elements like UAV efficiency, operational risk limitations, and the configuration of urban airspace. The risk-constrained A* (RC-A*) algorithm is proposed to solve this problem. The suggested method integrates a model for assessing operational risks that takes into account the UAV’s performance metrics, the density of ground workers, and strategies for risk mitigation. It includes a cost function that integrates distance costs and risk costs, a risk-based heuristic function, and a turning angle constraint to enhance search efficiency. Simulation trials grounded in a real-world context illustrate that the proposed algorithm is capable of devising secure trajectories for UAVs within urban three-dimensional environments. In the two groups of experiments, the operations risk values of paths are reduced by 67.34% and 57.98% compared to the standard A* algorithm.

Detection algorithm of ship key parts based on YOLOv11
WANG Yao, CAO Silei, LIU Minggang, WANG Yingying
2026, 52(11): 3969-3981. doi: 10.13700/j.bh.1001-5965.2024.0853
Abstract:

In contemporary and forthcoming maritime rights safeguarding and oceanic surveillance endeavors, it is critically important to investigate methods for swiftly and precisely acquiring data on essential components of vessels and adeptly analyzing real-time marine conditions. Despite the capability of deep learning-based object identification algorithms to identify various types of target information, there is a deficiency in research focused on algorithms for recognizing critical components of ships, and the examination of the relationship between multi-channel feature maps is inadequate. The feature utilization rate of key parts of ships with different scales is low, which may lead to the loss of important feature information and a decrease in detection accuracy. To address this issue, a ship key part detection method based on YOLOv11 is proposed to improve the accuracy and efficiency of key part detection. This manuscript initially performs a thorough examination of the critical and challenging issues in the present detection of essential ship components and engages in a comparative analysis with existing methods. Then, an improved YOLOv11 object detection network for ship key part detection is proposed based on actual problems. Ultimately, a dataset of key components for visible light ships has been established, and through the implementation of comparative and ablation experiments, it is thoroughly evidenced that the proposed algorithm developed in this study exhibits high precision and promptness in identifying ship key components, thereby offering viable technical assistance and a theoretical foundation for the meticulous differentiation of essential and critical parts across various ship types in intricate marine settings and the efficient surveillance of maritime vessel information.

End-to-end autonomous navigation method for drones based on AR-PPO
LI Hanwei, LI Xinkai, ZHANG Hai, JIANG Lingfeng, MENG Yue, ZHANG Hongli
2026, 52(11): 3982-3994. doi: 10.13700/j.bh.1001-5965.2024.0582
Abstract:

We propose an end-to-end autonomous navigation method for drones based on action masking-temporal-proximal policy optimization (AR-PPO) to address problems with deep reinforcement learning (DRL) algorithms for drone navigation, such as partial observability of the environment and high training difficulty. First, we introduce a gated recurrent network to process historical information, allowing the policy network to incorporate historical baseline information to tackle the problem of partial observability in complex environments. Second, we design an action masking method suitable for continuous action space reinforcement learning tasks, which excludes the interference of illegal actions and accelerates the model training speed. Then, we establish an actor-critic framework for a three-dimensional continuous action space DRL end-to-end model, which directly outputs control actions based on sensor perception results. Finally, we conduct comparative experiments with mainstream DRL methods. With an over 85% success rate in four generalized test contexts, the AR-PPO approach has the fastest training convergence speed and the largest rewards. Drones can safely perform obstacle avoidance navigation in real machine experiments. The results show that the proposed method effectively enables precise obstacle avoidance and safe navigation for drones.

Adaptive repulsive potential field for UAV trajectory planning
GUO Jing, XIAN Yong, REN Leliang, LI Shaopeng
2026, 52(11): 3995-4008. doi: 10.13700/j.bh.1001-5965.2024.0569
Abstract:

Aiming at the difficulty of potential field function design and the poor scene adaptability of the traditional artificial potential field (APF) method in multi-obstacle complex scenes, a three-dimensional self-adaptive repulsive potential field trajectory planning (TAPF) method based on the deep deterministic policy gradient (DDPG) algorithm is proposed. The adaptive repulsive potential field is used for trajectory planning. Firstly, a reinforcement learning-based repulsive potential field learning intelligence is established, which achieves “end-to-end” repulsive force output and improves the adaptability of the repulsive field. Secondly, dynamic triangles are designed to describe the range of the repulsive field, which effectively solves the problem of redundant paths in the traditional circular repulsive field guidance. Lastly, according to the characteristics of UAV trajectory planning, the state space, action space, and action space are designed for the TAPF method, and the proposed algorithm is based on the characteristics of unmanned aerial vehicle (UAV) trajectory planning. The algorithm’s state space, action space, and reward function are created. According to simulation results, the proposed method has superior robustness to complicated scene adaptation and plans shorter courses with a better planning effect.

Lateral directional control of aircraft based on distributed electric propulsion concept
WANG Bin, YANG Youxu, FU Junquan, DUAN Yiwei, SHEN Junhui, WU Yifei
2026, 52(11): 4009-4018. doi: 10.13700/j.bh.1001-5965.2024.0670
Abstract:

Distributed electric propulsion (DEP) aircraft used multiple motor-propeller system distributed on the body to provide thrust and this configuration offers a new strategy for lateral control of the aircraft. The aerodynamic effects of distributed slipstream were taken into account when developing the flight dynamics model for DEP aircraft. The traditional proportional-integral-derivative (PID) control algorithm was used to adjust the power output of the DEP system, controlling the aircraft’s lateral movement to improve flight stability and controllability. The results show that during yaw motion, the aircraft uses differential thrust control effectively tracking yaw angle commands, achieving a 30° yaw and reaching excellent con-trol performance. The airplane maintains a stable flight state during roll motion by using the slipstream lift effect of various thrust propellers to achieve a 30° roll.

Visual language model based on improved Transformer
JIN Zehui, CHEN Kai, TANG Dunbing, FANG Guoyu, WEN Feixia
2026, 52(11): 4019-4028. doi: 10.13700/j.bh.1001-5965.2024.0656
Abstract:

Natural language processing and computer vision are combined in the cross-modal job of image captioning. The model is required to generate an appropriate caption for the given image. To address this challenge, we proposed residual gated Transformer (RGFormer) as an enhancement of the Transformer architecture. The model, based on CLIP and RGFormer, CRM, is then proposed for image captioning. CRM utilizes the encoder of CLIP to extract image features as a prefix to the caption. RGFormer, a lightweight mapping network, projects the prefix into language space before feeding it into GPT-2 to produce captions. CLIP was trained on an extensive dataset comprising image-text pairs, which contains rich visual and semantic information and is exceptionally well-suited for vision-language tasks. The core idea of CRM is to reduce the disparity between visual and textual representations by using RGFormer to accomplish the cross-modal task. CRM could generate meaningful captions for large-scale datasets in a short training time without additional annotations or pre-training. Quantitative evaluation experiments show that CRM achieves results comparable to some advanced models on the COCO Caption dataset more efficiently.

Workspace stability evaluation of multi-engine suspension system based on EWM-TOPSIS
LI Wei, ZHAO Zhigang, ZHAO Xiangtang, LI Zixuan, GANG Zheng
2026, 52(11): 4029-4039. doi: 10.13700/j.bh.1001-5965.2024.0546
Abstract:

In response to the challenge of assessing the stability of a multi-robot suspension system’s workspace, we propose a stability evaluation method based on entropy weight method (EWM) and technique for order preference by similarity to an ideal solution (TOPSIS). Additionally, article establish a multi-dimensional workspace to characterize the distribution of stability within the workspace. Firstly, by analyzing the kinematics and dynamics characteristics of the system, article introduce performance indices for position, attitude, and flexure force uniformity that impact the system’s stability. Secondly, the article employs EWM-AHP to calculate composite weights for each performance index in order to determine their relative importance. Subsequently, TOPSIS is utilized to evaluate the stability performance of each position in the workspace. Finally, simulation computations are performed to identify the static balancing workspace of a fixed-end multi-machine suspension system and evaluate the stability of different positions within this space over various stability performance intervals. The simulation findings show that the multi-dimensional working space decreases toward its center and upper half as stability performance increases, indicating improved overall stability in these areas for the working space of the multi-machine suspension system. This analysis provides a theoretical foundation for optimizing configurations in multi-robot suspension systems.

Lightweight multi-scale feature fusion enhancement algorithm for spatial non-cooperative small target detection
LIU Zhenjiang, ZHANG Huijuan, JI Miaoxin, LU Tao, ZHANG Chi, YU Yuanjin
2026, 52(11): 4040-4050. doi: 10.13700/j.bh.1001-5965.2024.0509
Abstract:

In deep space environments, the feature of the non-cooperative small targets cannot be fully extracted, so the detection accuracy of uncooperative small targets will decrease. Furthermore, the computational resources are limited for space applications. To solve the above-mentioned problems, a lightweight multi-scale feature fusion, enhancement and context-aware spatial target detection algorithm is proposed. Multi-scale characteristics of spatially non-cooperative targets are extracted using a lightweight multi-feature downsampling module, which may improve the capacity to obtain fine-grained features. A fast dual spatial pyramid pooling module is developed to enhance the extraction of global contextual features, and the locally salient and global features are aggregated to improve the ability of feature generalization. A cross-dimension triple feature enhancement module (CTFEM) is designed, which could model the global context in term of different dimensional information. Therefore, the CTFEM module could further enhance the feature information of small targets and suppress the confused background information. The VisDrone2019 dataset and a self-built spatial situational awareness dataset (SSAD) are used to validate the suggested method in this research. The experimental findings show that the proposed method decreases the model parameters by 19.38% when compared to the benchmark model YOLOv8n. On the SSAD dataset, the mAP50 and mAP50-95 of the proposed method are improved by 0.2% and 1.2%, respectively. On the VisDrone2019 dataset, the mAP50 and mAP50-95 of the proposed method are improved by 2.0% and 1.4%, respectively.

Multi-core contextual feature-guided algorithm for trusted detection of UAV aerial images
ZHANG Zhihao, DU Lixia, HAO Ziwei, HOU Yue
2026, 52(11): 4051-4062. doi: 10.13700/j.bh.1001-5965.2024.0548
Abstract:

A multi-core contextual feature trusted detection network (MCCF) is developed for unmanned aerial vehicle (UAV) tiny target detection in order to address the issue of low detection accuracy caused by uneven target scale distribution in UAV aerial photos under complicated background. Firstly, a multicore contextual feature guidance module is designed in the backbone network part, which utilises multi-scale convolutional kernel and global contextual attention in order to achieve the fusion of feature information of local details and global structure, so as to effectively suppress the background noise. Secondly, in order to minimize the loss of spatial position information of small targets during the sampling process, a learnable position encoder is suggested that dynamically modifies the position representation of target characteristics. Finally, in order to effectively identify tiny targets in complex regions, this paper proposes a multi-scale detail fusion module, which fuses local spatial position information and global structural semantic information in order to effectively capture fine-grained features of targets at different scales. Finally, the experimental results on the VisDrone2019-DET dataset and the Drone dataset show that the mAP0.5 of the MCCF model is improved by 2.1% and 6.8%, respectively, compared to the RT-DETR model, which effectively improves the problems of target leakage and misdetection in UAV aerial images.

BN-based comprehensive evaluation model for display interface usability in civil aircraft cockpit
XIAO Yao, CHEN Dengkai, ZHOU Yao, XIAO Jianghao, HOU Yuhong
2026, 52(11): 4063-4074. doi: 10.13700/j.bh.1001-5965.2024.0627
Abstract:

Optimizing the design interface is aided by research on the usability of display interfaces in civil aircraft cockpits. Aiming at the uncertainty of usability evaluation of civil aircraft cockpit display interface, this paper proposes a comprehensive usability evaluation model based on the Bayesian network (BN). Firstly, a Bayesian network is established based on the usability evaluation indicator system obtained from the executive process-interactive control model (EPIC). Both the experts' triangle fuzzy number evaluation and the quantitative root nodes' weighted evaluation of expert eye movement measures served as prior probabilities. Secondly, in order to reduce the number of experts’ decision-making, the Noisy-OR model was used to fill the conditional probability table of the Bayesian network, and the model was built using GeNIe software. A primary flight display interface is taken as an example for instance verification. The results show that the causal reasoning results of the model are consistent with evaluation results of the system usability scale (SUS) scale. The evaluation grade of the interface is "good", which verifies the effectiveness of the model, and proposes improvement measures for sensitive factors obtained by the model sensitivity analysis.

A real-time visual SLAM algorithm for complex dynamic scenarios
ZHAO Boting, LIU Jun, CHAI Hongxu, ZHANG Jianye, WU Ruibin
2026, 52(11): 4075-4089. doi: 10.13700/j.bh.1001-5965.2024.0658
Abstract:

Simultaneous localization and mapping (SLAM) systems frequently have a number of issues in general or sophisticated dynamic scenarios with dynamic objects, including strong assumptions, inadequate positioning precision, or even failure. To solve these problems, a real-time visual SLAM algorithm for complex dynamic scenarios is proposed, which is built on ORB-SLAM3. The algorithm deploys YOLOX-s based on accelerated model inference to obtain semantic prior information to process dynamic objects, and the Kalman filter and Hungarian algorithm are combined to compensate for missing object detection. A two-category dynamic scenario processing strategy is proposed based on the grid-based motion statistics (GMS) feature matching algorithm to rapidly screen reliable matching points, which separates the scenarios into two categories based on the image matching accuracy, in order to enhance the positioning accuracy and generalization ability of the algorithm. For low dynamic frames, the GMS algorithm is used to quickly filter out dynamic features, so as to avoid a large number of matrix operations and improve operational efficiency. For high dynamic frames, the Chi-square test is utilized to identify motion outliers, and a dynamic feature culling strategy is proposed to cull dynamic points according to the adaptive threshold and keep the tracking stability. The experimental results on standard datasets and real scenarios show that the algorithm effectively addresses the real-time localization challenges, demonstrating its superior accuracy, robustness and real-time performance in complex indoor and outdoor dynamic scenarios.

Flight task recognition and action segmentation based on SVM
LI Bowen, LEI Xiaoyong
2026, 52(11): 4090-4097. doi: 10.13700/j.bh.1001-5965.2024.0635
Abstract:

The quality assessment of civil aviation flight training is gradually developing in the direction of quantification and intelligence. A support vector machine (SVM)-based approach for civil aviation flight task recognition and action division is suggested in order to satisfy the new requirement of flight task recognition. Firstly, civil aviation flight parameters are collected and preprocessed for subsequent flight task identification and action division. Then, the support vector machine model is used to identify the civil aviation flight tasks. The traditional method uses a flight track image for recognition, but the accuracy of take-off and landing recognition is low due to the lack of time sequence features. This issue is resolved by converting xyz coordinates to RGB pictures, adding numerical channels to represent the data's absolute value, and introducing time series characteristics. Compared with the experimental results of traditional track image recognition, RGB image recognition, and the combination of RGB image and numerical channel recognition, it is found that the recognition accuracy is increased by 43% by using the method of combining RGB image and numerical channel, which can effectively improve the recognition accuracy of turning around landmarks and rectangular routes. In addition, according to the flight task identification results, each task is divided into nine types of flight actions: take-off, landing, climb, descent, cruise, climb turn, descent turn, cruise turn and hover, which provides support for the subsequent flight training quality assessment research.

Target tracking algorithm based on IGWO-STCKF
WANG Wei, GONG Shuli, LI Xiaoming
2026, 52(11): 4098-4106. doi: 10.13700/j.bh.1001-5965.2024.0623
Abstract:

Aiming at the problems the cubature Kalman filter(CKF) algorithm needs to predetermine the filtering noise and that the tracking accuracy decreases when the target is maneuvering, a strong tracking cubature Kalman filter algorithm based on improved grey wolf optimization (IGWO-STCKF) is proposed. Firstly, the process of CKF is introduced. Secondly, the grey wolf optimization is used to achieve the adaptability of the process noise covariance. The good point set theory is used to obtain the initial population, embedding the nonlinear adjustment strategy of control parameters, the out-of-bounds processing strategy, and the dynamic weight adjustment strategy, in order to address the shortcomings of the grey wolf optimization (GWO) algorithm, which include its slow search speed and ease of achieving a local optimum. Meanwhile, the fading factor, which takes reference from a strong tracking filter (STF), is used to adjust the covariance matrix during the measurement update period and adjust the filtering gain in real time. Finally, IGWO-STCKF algorithm is formed. IGWO-STCKF, IGWO-CKF, STCKF, IMM-CKF and CKF algorithms are applied to a two-dimensional maneuvering target tracking simulation experiment with dynamic noise, and the tracking trajectory diagram, algorithm execution time, the root mean square error curves of position and velocity, and cumulative root mean square error are analyzed. The results show that IGWO-STCKF algorithm has high filtering accuracy, strong robustness, and better tracking effect.

A bi-level programming method for urban low-altitude travel public air route network driven by multi-source data
WANG Weiwen, BAO Jie, CHEN Fangxiao, ZHOU Jun, YANG Zhao
2026, 52(11): 4107-4121. doi: 10.13700/j.bh.1001-5965.2025.0288
Abstract:

Urban low-altitude travel is the core of high-quality development in the low-altitude economy and an important link for enabling low-altitude empowerment and economic radiation. For urban low-altitude public flight routes, this study suggests a multi-source data-driven dual-layer planning approach that creates a feedback optimization mechanism between low-altitude demand allocation and flight route network planning. A road impedance function calibration method tailored for low-altitude travel is presented, and the multi-objective simulated annealing-genetic algorithm (MOSA-GA) hybrid evolutionary algorithm is designed for solution search. Finally, taking a typical metropolitan area as a case study, multi-source urban data such as airspace, buildings, land use, pedestrian flow, and noise distribution are extracted, and the proposed dual-layer planning method is validated through comparison. The experimental findings demonstrate that the MOSA-GA algorithm successfully strikes a compromise between the optimality of the goal functions and the diversity of the solution set, greatly enhancing the low-altitude route network’s operational efficiency. The solution results yield multiple Pareto-optimal solutions that are mutually non-dominated, with significant trade-offs between multiple objectives, which impact traffic flow allocation. Additionally, sensitivity analysis reveals that safety intervals and minimum segment lengths significantly affect the network optimization outcomes. The research findings are expected to guide the deployment of urban low-altitude public flight route networks and improve the level of low-altitude travel services.

UAV perspective target detection algorithm based on lightweight YOLOv8n
SONG Junyu, PANG Yan
2026, 52(11): 4122-4134. doi: 10.13700/j.bh.1001-5965.2025.0417
Abstract:

In order to ensure the accuracy of target detection within limited resources, a lightweight YOLOv8n target detection algorithm YOLOv8n-WMCBS from the perspective of an unmanned aerial vehicle (UAV) was proposed. Firstly, Wise-IoU v3 is used as the overall loss function of the network, and the low-quality prediction frame is dynamically focused. Secondly, the MobileNetV3 network and the coordinate attention (CA) backbone network were introduced to reduce the number of parameters and the computational complexity. In order to preserve the specific characteristics and position data of the small target and lessen the model’s computational load, the big target detection layer is eliminated and the small target detection layer and detection head are added concurrently. Then, the residual ghost convolution (RGConv) was designed to further enhance the feature richness of the backbone network, and was combined with the bidirectional feature pyramid network (BiFPN) module as a new neck network module to enhance the dynamic feature fusion of the neck network. Finally, in order to fully preserve the high-resolution small target information, an enhanced convolution-space-to-depth (Conv-SPD) is suggested to replace the initial step-down down sampling. Experimental results show that in the VisDrone 2019 dataset, while the number of parameters and the model size are reduced by 40.5% and 35.5%, the mAP0.5 and mAP0.5:0.95 are increased by 4.6% and 2.9%, respectively, compared with the original algorithm, which has stronger generalization ability and is more suitable for resource-constrained equipment such as UAVs.