Abstract:Addressing the severe challenges of multipath fading, Doppler shift, and complex interference faced by Unmanned Aerial Vehicles in high-speed mobile scenarios, this study proposes and implements a set of key physical-layer technologies for a high-performance data link. Core innovations include: a rapid signal acquisition technique based on the joint detection of multi-segment independently modulated Pseudorandom Noise codes, effectively resolving the significant Doppler shift (up to ±54 kHz) and timing synchronization challenges caused by high-speed mobility, achieving an acquisition success rate exceeding 99%; a cyclic sliding correlation channel estimation method combined with a Single-Carrier Frequency-Domain Equalization with Iterative Decision Feedback architecture, significantly enhancing the robustness of carrier tracking and signal demodulation in fast-fading multipath channels; an overlapping window frequency-domain notch filtering technique integrated with adaptive spectrum sensing, enabling effective suppression of narrowband, mid-band, and full-band interference. Theoretical analysis, formula derivation, and simulation verification demonstrate that the proposed scheme achieves an approximate 4 dB performance gain compared to traditional 1-bit differential demodulation across various complex channel environments (e.g., suburban, urban), substantially improving communication reliability for high-speed mobile platforms.

Abstract:Aircraft icing detection is crucial for ensuring flight safety. Ice thickness measurement can provide a quantitative assessment of ice accretion on different key parts of the aircraft. This paper utilizes the principle that ultrasonic waves generate pulse echoes when propagating across media. By collecting the reflected signals of ultrasonic pulse echoes, the time of ultrasonic wave transmission in the ice layer is determined to sense the thickness of the ice layer covering the measured surface. Based on the measurement principle of ultrasonic waves and the actual material of the aircraft skin, an ultrasonic generation system device and a semiconductor refrigeration icing platform were built. Through experimental measurements on simulated ice layers with thicknesses ranging from 1 to 17 mm on the aircraft skin, it was found that the ultrasonic wave measurement results for thicker ice layers (≥ 2 mm) were accurate, with the measured values matching the actual simulated ice layer thicknesses, and the precision could reach±0.5 mm. However, for thinner ice layers, the measurement accuracy was poor, and in some cases, detection was even impossible. The research in this paper provides experimental evidence for the engineering application of in-situ ice thickness measurement based on ultrasonic pulse waves on aircraft and lays the foundation for the further development of excellent ultrasonic icing detection sensors.

Abstract:In light of the reform of the traditional flight test mode, there is a proposal to construct the intelligent automation capability in the “Digital empowerment flight test”. To address the need for high-definition video recording of the target′s key motion process in the near-field runway segment flight test, a target positioning and tracking method based on the ground-based PTZ. This method calculates the relative azimuth information of the target in real time using the GPS/Beidou differential calculation model. Based on this information, the PTZ is guided to rotate and adjust the focal length of the high-definition camera, enabling automatic capture of the target′s field of view. Real-time detection and recognition of the target are then carried out using the YOLOv5s object detection algorithm, which has been improved with global attention mechanism, Dynamic head, and Wise IoU. By incorporating techniques such as median flow, Kalman filtering, and IoU threshold detection, long-term real-time tracking of the target is achieved. The experimental results demonstrate that the proposed method achieves a mean average precision of 76.3% for all categories, with a processing frame rate of 20 fps. This performance effectively supports the requirements of practical flight test engineering applications.

Abstract:For the challenges in the ground-based indoor field test and verification of the entire process of high-precision measurement and localization of time difference of arrival(TDOA) and frequency difference of arrival (FDOA) by multi-observation station formations in dynamic scenarios, a semi-physical testing method for multi-observation station localization system of TDOA and FDOA is proposed. The multi-stations localization system of TDOA and FDOA works collaboratively with dynamic scene simulators, agile signal generators, atomic clocks, system calibrations and other auxiliary equipments under a unifield time-frequency reference and control instructions, achieving a full-scale realistic simulation and test verification of the TDOA and FDOA variability environment and inter-statation time synchronization and frequency phase consistency constraints faced by the implemented multi-station system. The test result show that the TDOA measurement accuracy is better than 10 ns(r.m.s), the FDOA measurement accuracy is better than 1 Hz(r.m.s), and within a range of 900 km from the radiation source to the observation stations,the localization CEP accuracy is better than 100 meters. Compared with traditional field implemention tests, the test time cost is reduced by approximately 80%, and the test cost is reduced by approximately 70%, thereby greately improving the system testing efficiency and coverage, and releasing the techincal and engineering risks of the multi-station formation collaborative localization system in the ground indoor field measurement environment.

Abstract:To obtain more abundant data on the inlet pressure flow field of aero-engine, this paper adopts the concept of probe fusion and designs a composite pressure probe capable of simultaneously measuring single-point total pressure (low-frequency/highfrequency) and static pressure (low-frequency) in the flowfield. The probe has been developed, installed on an aircraft for flight testing, and validated. The results demonstrate: the newly equipped pressure probes exhibit a blockage ratio of 0.56%, with structural strength meeting installation requirements. Flight tests have successfully captured synchronized measurements of the steady-state total-pressure field, steady-state static-pressure field, and dynamic total-pressure field at the engine inlet, providing abundant and accurate flowfield data; compared to directly using dynamic pressure sensors, employing pressure tubes introduces a slight time lag in the measured data. Therefore,to minimize the impact of pipeline cavity effects and ensure the accuracy of measurement data in inlet/engine compatibility flight tests, on the one hand, the pressure routes for the inlet flow field measurement system be optimized for the shortest distance. On the other hand, the steady-state test points should be maintained for at least 3 seconds or more.

Abstract:In order to solve the problem of insufficient research on the aerodynamic characteristics of the transonic probe at subsonic flow in the existing research, the structure design of the transonic wedge-shaped four-hole probe was carried out firstly, and the angle between the two inclined planes and the centerline of the designed probe was 17°, the diameter of the bleeding air hole is 1.1 mm. The flow field and structural strength of the probe are simulated by the fluid-structure interaction method, and it is found that under the supersonic condition, a stable appendage oblique shock wave is formed in front of the probe tip, and the maximum stress of the probe in the flow field is 142.86 MPa, the safety factor is 2.67. Finally, the probes were calibrated pneumatically under subsonic and supersonic conditions, and it is found that the calibration curves were relatively regular, the angle sensitivity was high, and the orthogonality between the calibration points was good. Under subsonic conditions, when the pitch angle of the probe is positive and the absolute value of the yaw angle is ≥5°, the orthogonality of the curve is poor, and with the increase of Ma, the orthogonality of the characteristic curve is improved.

Abstract:Defect detection of aero-engine blades is a critical technological step to ensure flight safety. Traditional industrial borescope inspection methods rely heavily on manual expertise, which often leads to low efficiency, strong subjectivity, and missed detections of subtle defects. To address these issues, this paper proposes a lightweight and high-precision improved YOLOv11 model, specifically designed for real-time defect detection of engine blades. A high-quality industrial image dataset was constructed, comprising four typical defect types—bending, ablation, cracks, and material loss. Aiming at the characteristics of small targets, complex backgrounds, and multi-scales, a CFES backbone network is constructed to enhance the integration ability of semantic information and reduce the amount of computation. Specifically, ShuffleNetV2 is employed as the backbone network instead of the original one to alleviate computational overhead, while the BiFormer attention mechanism is integrated to strengthen the feature representation capability. Additionally, a Dynamic-DCNv3-based detection head is employed to improve the modeling capability for complex textures and small-sized objects. Experimental results show that the improved model achieves an mAP@0.5 of 85.0%, surpassing the baseline model, while reducing parameters to only 1.7 M. This demonstrates superior detection performance and adaptability for edge deployment. The model was successfully deployed on an industrial borescope platform equipped with the RK3588 chip, where the frame rate remained at approximately 30 frames per second, achieving efficient and stable automatic defect recognition. This study provides a practical solution for intelligent on-site inspection in aviation maintenance and offers technical support for the application of lightweight object detection models in industrial embedded scenarios.

Abstract:Aiming at the problems of poor compatibility with external data sources, low efficiency in field configuration iteration, and limited reliability of data transmission in existing aircraft test systems, an intelligent test system based on a multibus collaborative architecture is proposed. The system employs an FPGA as the core processor, supporting the reception, fusion, and transmission of data through three bus interfaces: RS422, 1553B, and Ethernet. A gigabit Ethernet communication module is designed to enable real-time configuration command delivery and online updates of program memory, along with a functional module for processing the UDT protocol. Experimental and simulation results demonstrate that the system can collect and store RS422 data and 1553B bus data, achieving multi-bus data collaborative recording. The online upgrade function operates successfully in temperature environments of -40℃~100℃, with self-tests showing upgrade completion times consistently under 1 second. In simulated data input scenarios, using the UDT protocol achieves a data packet delivery accuracy rate of 99.8%, a significant improvement over the 87.5% rate of the UDP protocol, thereby enhancing data transmission reliability.

Abstract:In flight test, real-time monitoring is an important part to ensure flight safety. To address the challenges of parallel processing and comprehensive monitoring of multiple telemetry data streams in complex flight test scenarios like multi-plane collaboration characterized by large data volumes, strong real-time requirements, and high risk levels, an overall architecture for flight monitoring oriented towards multiple telemetry data streams was constructed. An adaptive architecture-based dynamic planning and resource scheduling strategy for flight missions was designed, and a parallel processing method for multitelemetry data streams based on a stream computing framework was proposed. Additionally, a fused publish/subscribe mechanism for multi-aircraft telemetry parameter in complex flight test scenarios is established. Application results demonstrate that this system can achieve rapid reconstruction of monitoring tasks and resource scheduling, as well as parallel processing and on-demand distribution of multi-telemetry data streams, effectively meeting the real-time data processing and safety monitoring requirements for complex flight test subjects.

Abstract:To accurately obtain the dynamic loads of the helicopter tail rotor within the flight envelope and address the challenges of low reliability in traditional slip rings and wired telemetry technologies, wiring difficulties, and insufficient synchronization accuracy, this paper developed a four-channel tail rotor load measurement system integrating wireless transmission and high-precision clock synchronization technologies. The system consists of a wireless load collector deployed on the rotating components of the tail rotor and a load receiver in the cabin, transmitting data via an RF link. It employs a global time reference based on inter-range instrumentation group time code B signal synchronization and an optimized master-slave clock synchronization algorithm, combined with dynamic timestamp alignment and strain measurement temperature compensation techniques, aiming to achieve microsecond-level data synchronization. Experimental results demonstrate that the constructed four-channel measurement system achieved an overall average synchronization accuracy of 1 060.5 ns, with the best channel reaching an average synchronization accuracy of 799 ns. Each strain measurement channel exhibited excellent linearity and measurement accuracy within the range of ±10 000 με, with the root mean square error of the average code values all below 3.9, meeting the requirements for flight tests.

Abstract:This paper focuses on the wireless transformation of airborne flight test systems, which can solve the problem such as complex cabling, difficult installation and removal, time-consuming setup, and maintenance difficulties encountered during the development of flight testing. The flight test system equipment includes various sensors and storage devices. Leveraging the advantages of SparkLink wireless communication technology—such as low energy, low latency, high throughput, and high reliability—customized compact short-range wireless communication modules were designed. These modules were integrated with sensors, storage devices and customized wireless module, enabling analog-to-digital converted sensor data to be transmitted over the air to the storage units. This achieves wireless recording of flight test data and a wireless flight test system. Test results under laboratory conditions demonstrate that the customized wireless modules can achieve data transmission rates up to 4 Mbps with a transmission delay below 3.5 ms, meeting the performance requirements for airborne flight test data recording.

Abstract:As the core power component of aircraft, the operational reliability of aero-engines is directly related to flight safety and efficiency, and fault diagnosis of intershaft bearings is a key measure to ensure their stable operation. Aiming at the fault diagnosis problem of intershaft bearings in aero-engines, this study first analyzes the limitations of existing 1D-CNN and 1D-Transformer methods: the self-attention mechanism is susceptible to high-frequency noise and redundant information in raw vibration signals, which weakens the ability to focus on critical fault features; meanwhile, the pure Transformer architecture shows insufficient capability in capturing subtle local features. To address these issues, a Multi-Scale Time-Frequency Synergy Transformer based fault diagnosis method is proposed, which integrates multi-scale time-frequency feature extraction with the global modeling capability of the Transformer, enabling collaborative capturing of both subtle local features and global correlation features of vibration signals. Experimental results indicate that in Gaussian white noise environments (SNR from -4 dB to 4 dB), the proposed method exhibits excellent fault diagnosis performance for aero-engine intershaft bearings: both diagnostic accuracy and F1-Score are optimal, reaching 96.04% under strong noise (-4 dB) and 99.84% under weak noise (4 dB), with noise-resistance stability superior to five benchmark methods. On the CWRU benchmark dataset, in both noise-free and noisy scenarios, it can stably identify different fault severities (including slight inner-race faults), achieving 99.01% accuracy under strong noise (-4 dB) and 99.78% under weak noise (4 dB), thereby demonstrating its strong generalization capability. In conclusion, the proposed MSTFS-Transformer effectively alleviates the insufficient feature-focusing and weak local feature-capturing problems under noise interference, providing an efficient and robust solution for aero-engine intershaft bearing fault diagnosis. Its strong noise immunity and accurate identification capability meet the demands of complex vibration environments in practical engineering, and offer solid technical support for improving fault-monitoring reliability.

Abstract:To address the challenge of quantitatively evaluating the structural reliability of probe in high-temperature flow fields of aero-engines under multi-load coupling effects, a fatigue-creep coupled failure analysis method was developed specifically for this probe. Based on the one-way fluid-thermal-solid coupling method, the structural strength of the probe was evaluated. The results indicated a maximum temperature of approximately 850℃, a maximum static stress of 209.4 MPa, and a maximum random vibration stress of 44.6 MPa. The Basquin equation combined with the Goodman correction model was applied to assess high-cycle fatigue damage, yielding a fatigue cycle number of 1.23×1018. The Larson-Miller equation was utilized to calculate a creep cycle number of 1.03×109. By employing the linear cumulative damage model, the service life under the coupling of multiple damage mechanisms was analyzed. It was determined that with a safety factor of 2, the service life is ≥472.1 h, meeting the design requirements. In-engine verification demonstrates that the probe has safely operated for over 50 h. This method enables the quantitative assessment of high-temperature probe failure, providing a technical foundation for the future design of high-temperature probe in aero-engines.

Abstract:To address the critical challenge of aircraft icing in complex atmospheric environments, this study aims to develop infrared-responsive photothermal materials for efficient de-icing applications. A Cu3BiS3/Bi2S3 heterojunction nanorod structure with localized surface plasmon resonance (LSPR) characteristics was constructed via an ion exchange strategy, forming a stable p-n interface. The structural, optical, and photothermal properties of the material were systematically characterized. Results reveal a distinct LSPR absorption peak near 980 nm and a strong absorption tail extending into the near-infrared region, with considerable absorption retained at 808 nm. Under infrared laser irradiation, the heterostructure exhibits rapid surface temperature elevation up to 70℃ within 10 minutes, significantly outperforming pure Bi2S3. Furthermore, femtosecond transient absorption spectroscopy reveals that LSPR-excited hot carriers undergo efficient interfacial separation and extended lifetimes at the heterojunction, enhancing nonradiative energy dissipation and overall photothermal conversion. This work provides a promising strategy and mechanistic insight for the development of high-efficiency, low-power infrared photothermal de-icing materials for next-generation aerospace applications.

Abstract:Aiming at the problem of GNSS signal loss caused by complex airspace environments or highly dynamic maneuvers during flight testing, an algorithm is proposed that corrects INS velocity to GNSS velocity based on the lever arm effect. Unlike standard lever arm velocity compensation models, this paper takes into account angular acceleration and centripetal acceleration during lever arm rotation to adapt to the highly dynamic maneuvers of aircraft. The Euler integration of the corrected GNSS velocity in the ECEF coordinate system is implemented, and cumulative errors are reduced through simultaneous forward and backward integration. The effectiveness of the algorithm is verified using real flight test data. Experimental results simulating 10 seconds of GNSS signal loss show that when the aircraft is in maneuvering states such as climbing or rolling, the cumulative integration error does not diverge, and the algorithm′s calculated results maintain an error level relative to GNSS ranging from centimeters to meters.

Abstract:Aiming at the problems that the existing decision tree methods rely heavily on prior knowledge in setting threshold values and the neural network methods in the field of communication signal recognition have large model sizes and high parameter counts, this paper proposes a lightweight neural network modulation recognition method integrating decision trees. This method introduces the idea of decision trees to analyze the confusion matrix of the dataset, divides the dataset into different subclasses based on the characteristics of different signal categories, and uses lightweight convolutional neural networks for hierarchical recognition. To achieve targeted recognition for each subclass in the hierarchical recognition process, data cleaning and feature extraction are employed to obtain the unique signal features of each subclass. Experimental results show that on the public dataset RML2016.10a, the overall recognition rate of this method reaches 90.03% within the signal-to-noise ratio range of 0 to +18 dB, which is 7.49% higher than the highest recognition rate of the comparison models. When the signal-to-noise ratio is 18 dB, the recognition rate reaches 95.03%; and the model parameter count is only 86 342, which is 96.85% lower than that of models with the same accuracy.

Abstract:Addressing the issues of missed and false detections of pallets in real factory environments, often caused by factors such as insufficient lighting and numerous obstacles, a tray detection method based on an improved YOLOv8n is proposed. Firstly, the Bi-Level Routing Attention (BRA) sparse attention module combined with Transformer is incorporated into the backbone network feature extraction phase of the YOLOv8n model, to reduce the interference from obstacle occlusion on pallet detection. Secondly, the Shape-IoU loss function is introduced, further enhancing the model′s ability to recognize pallets in conditions of insufficient lighting and severe background interference. Finally, the feature fusion network of YOLOv8n is reconstructed using the GSConv-based Slim-neck structure, achieving a lightweight neck network. Experimental results indicate that the improved algorithm achieves a mean Average Precision (mAP) of 89.6% on the test set, representing a 2.8% improvement compared to the original model. The missed detection rate and false detection rate decrease by 2% and 2.2%, respectively. This effectively mitigates the problems of missed and false detections of pallets in situations of insufficient lighting and obstacle occlusion. Additionally, with a detection frame rate of 312.5 fps, the method enables rapid and accurate pallet detection and recognition, making it suitable for deployment on smart forklifts to enhance operational efficiency and elevate warehouse intelligence levels.

Abstract:Real-time high-precision segmentation of complex urban street scenes is crucial for autonomous driving. Aiming at the problems that existing real-time semantic segmentation networks have insufficient capture of spatial information and detailed features in high-resolution branches, as well as inefficient fusion of high and low-resolution features leading to information loss, which restricts the improvement of segmentation accuracy, this paper proposes a real-time semantic segmentation network based on multi-scale partial dilated convolution and boundary collaborative dual attention guided fusion(MPDANet).First, a Multi-Scale Partial Dilated Convolution Module (MSPDC) is designed. By using partial dilated convolutions with parallel ladder-type dilation rates, it efficiently captures detailed features andspatial information of high-resolution branches from different scales, addressing the problem of insufficient information capture.Second, an Attention-Guided Feature Pyramid Module (AFPM) is constructed. It extracts multi-scale semantic information from low-resolution branches through an asymmetric pooling layer and further enhances the semantic information by combining a pixel attention mechanism.Finally, a Boundary Collaborative Dual Attention Fusion Module (BCDAF) is proposed. It screens key semantic and spatial information through parallel channel-spatial attention, suppresses information loss caused by cross-resolution feature fusion, and introduces boundary attention to improve the segmentation effect of target boundaries.On the Cityscapes validation set, the proposed network achieves 78.6% mIoU at a speed of 295 fps; on the CamVid test set, it achieves 77.4% mIoU at a speed of 454 fps. Experimental results show that the network proposed in this paper achieves high-precision segmentation of complex urban street scenes while maintaining real-time performance.

Abstract:Aiming at the problems of modeling difficulty and weak model generalization ability in complex assemblies, a position prediction method fusing the Improved Dung Beetle Algorithm (GTDBO) and Perceiver model is proposed. First, the ideal assembly is built and the assembly features under six-dimensional perturbation are collected, and a coupled dataset is constructed by interpolation. After that, the Perceiver model learns the nonlinear mapping relationship between feature deviation and positional deviation and optimizes its key hyperparameters with the help of GTDBO. The algorithm combines game-theoretic equilibrium control, adaptive angular perturbation and dynamic foraging strategies. Experiments on the CEC2017 test set show that the algorithm outperforms the comparison algorithms in terms of convergence speed and solution quality. Finally, the model is compared with Jacobian′s method, BP networks, SVR and the original Perceiver model in predictive control experiments on two types of artifacts. The results show that GTDBO-Perceiver achieves MAEs of 6.72×10-2 and 9.96×10-2 on the test set, respectively. Its ability to converge the deviation to the tolerance range within a finite number of control times and to achieve a balanced distribution of errors in defective piecewise scenarios demonstrates a good generalization capability.

Abstract:To address the issues of low accuracy and insufficient generalization ability of models in complex scenarios for current transformer appearance defect detection, this paper leverages the residual structure′s merits to improve YOLOv11n with three modules. Firstly, an inverted residual attention mechanism, iEMA, is designed. It can effectively utilize the long-distance dependency and aims to improve the accuracy of transformer defect detection. Secondly, by leveraging the advantages of depthwise separable convolution in multi-scale feature extraction and the characteristics of the residual structure, an MSCB structure is designed to enhance the feature extraction and fusion capabilities of the model. Since, to address missed detections due to insufficient contextual utilization by YOLOv11′s detection head, we design the MR-Detect head. It integrates grouped convolution and residual concepts, offering rich feature representations for classification. Finally, the non-maximum suppression algorithm is combined with Inner_MPDIoU to address the limitations of traditional loss functions in detecting irregular objects and objects with large size variations. Experimental results show that compared with YOLOv11n, the improved algorithm in this paper, while ensuring real-time detection, increases the mAP@0.5 by 5.9% and the recall rate by 2.8%. It has higher detection accuracy in complex transformer operating condition detection scenarios and can more effectively detect various types of defects.

Abstract:An information-rich and computationally efficient cost volume is crucial for high-precision and high-efficiency stereo matching. To construct such a cost volume and achieve accurate stereo matching, a lightweight network, Efficient-ACVNet, was proposed based on Fast-ACVNet to improve the efficiency of cost volume construction in stereo matching. First, a computationally less intensive 3D cost volume was used as cost volume attention. Inverse bottleneck residual blocks were employed to stack a symmetric hourglass structure for cost aggregation of the 3D cost volume, and multi-scale disparity channel attention modules was introduced to further enhance the aggregation effect. The aggregated 3D cost volume served as cost volume attention to construct and filter the information-redundant 4D cost volume, improving its information content and computational efficiency. Finally, pseudo-3D residual blocks were introduced and pseudo-3D downsampling modules was designed for cost aggregation of the 4D cost volume, further reducing network complexity. Experimental results showed that compared to the baseline method, the proposed algorithm reduced the endpoint error (EPE) by 9.375% on the SceneFlow dataset, decreased the outlier percentage (D1-fg) in the foreground region by 19% on the KITTI15 dataset, and reduced network runtime from 39 ms to 25 ms.

Abstract:There are many kinds of mushrooms, especially poisonous mushrooms, which are similar in shape and difficult to identify. There is an important practical need for efficient identification of mushroom species. In view of the problems of complex background, low recognition accuracy, large number of model parameters and difficult deployment on mobile terminals in existing mushroom image recognition methods, a mushroom image recognition method based on improved ConvNeXt model and knowledge distillation is proposed. Firstly, the pre-trained ConvNeXt weight file is applied to the mushroom recognition task through transfer learning, and the coordinate attention mechanism is introduced to construct the ConvNeXt_CA model, which effectively improves the fine-grained feature extraction ability of the model. Secondly, based on the knowledge distillation technology, the ConvNeXt_CA model is used as the teacher model and the ShuffleNet v2 model is used as the student model to construct a lightweight MushNet model. The overall efficiency of the edge deployment of the improved model is greatly improved. Finally, the relevant model comparison experiments are carried out. The results show that the accuracy of the proposed improved model reaches 94.89%, and the size of the MushNet model after knowledge distillation is about 1/21 of the original, which is better than other traditional models and lightweight models. The effectiveness and feasibility of the proposed mushroom image recognition method are proved.

Abstract:As critical components of aero-engines, turbine blades are prone to defects such as cracks, burns after long-term service, which can directly affect the safe and efficient operation of aircraft. To address the limitations of conventional machine vision or semantic segmentation methods in accurately segmenting blade defects under complex conditions, this paper proposes a Swin-DCUnet-based segmentation and assessment method for aero-engine turbine blade surface defects. The core of this approach is the semantic segmentation model Swin-DCUnet, which employs the Swin Transformer—capable of extracting multi-scale features—as the backbone feature extractor. The extracted features are fused through a dual-channel convolutional process, and a hybrid loss function is introduced to improve model convergence speed and segmentation accuracy. Furthermore, a defect severity grading method is developed by integrating the predicted segmentation results with quantitative analysis, providing a valuable reference for subsequent blade maintenance. A dedicated dataset and evaluation metrics are constructed, and ablation experiments are conducted. Results show that the proposed Swin-DCUnet achieves AR, AF1, mIoU, and Dice scores of 92.18%, 92.92%, 87.44%, and 47.85%, respectively, demonstrating its advancement, effectiveness, and practicality.

Abstract:To address the challenges of small object detection in UAV aerial imagery—such as dense distribution, scale variation, occlusion, and complex backgrounds—this paper proposes RAD-YOLO, an improved lightweight detection framework based on YOLOv11n. The model incorporates a RFM-FPN with RAU and SBA to enhance multi-scale feature integration. It also employs RFAConv in the backbone for adaptive spatial modeling, and introduces DDS-Soft-NMS strategy to reduce false suppression based on object scale. Experimental results show that RAD-YOLO improves mAP@0.5 and mAP@0.5:0.95 by 13.1% and 11.4% respectively on the VisDrone2019 dataset, achieving 0.561 precision and 0.411 recall. On AI-TOD and SODA-A datasets, mAP@0.5 improvements of 9.9% and 7.7% further demonstrate its robustness and strong generalization in complex aerial scenarios.