Abstract:The autonomous flight performance of unmanned aerial vehicles is significantly constrained by wind disturbances in practical applications such as logistics and inspection. Traditional filter-based methods rely on precise dynamic models and noise assumptions. When confronted with complex, time-varying wind disturbances, these approaches suffer from significant degradation in estimation performance due to unmodeled dynamics. In recent years, data-driven and hybrid methods have demonstrated tremendous potential. This paper aims to review research progress in UAV state estimation under wind disturbances, analyze the limitations of traditional Kalman filter-based methods in addressing wind disturbances, and explore the advantages of data-driven approaches in modeling and compensating for wind disturbances. It focuses on hybrid paradigms such as physically informed neural networks and model compensation learning, providing effective pathways for achieving highprecision, robust and generalized wind-resistant state estimation. Finally, this paper summarizes current technical challenges and outlines future research directions, particularly exploring the potential value of large language models in enhancing the intelligence and interpretability of state estimation systems.

Abstract:The automotive open system architecture (AUTOSAR) standard is provided in the form of large-scale textual specifications that are extensive and inherently ambiguous. To verify the compliance of in-vehicle software, consistency testing requires the extraction of configuration information from these documents to ensure conformity with the standard. Based on an in-depth analysis of AUTOSAR specifications and consistency test configurations, this paper proposes an automated method for generating configuration data. First, MinerU is integrated into a graph-based retrieval-augmented generation framework to construct a domain-specific entity-relation knowledge network through semantic enhancement, text refinement and structure-aware chunking strategies. Second, a few-shot chain-of-thought (Few-Shot CoT) prompting technique is employed to guide a large language model (LLM) in extracting test scenarios and the required configuration data from the domain knowledge network. Finally, a configuration item tree with attributes is generated according to configuration rules, and breadth-first search (BFS) and depth-first search (DFS) algorithms are applied to validate and complete the configuration data, resulting in ARXML configuration files compliant with the AUTOSAR standard. Experimental results on 15 AUTOSAR basic software modules, evaluated using 8 metrics and compared with traditional configuration data generation methods, show that the proposed approach achieves an F1 score of 97.51% in implicit dependency extraction, reduces the manual correction rate to 5% and increases the number of generated test scenarios by 3%. The configurations generated by the proposed method not only assist testers in identifying hard-to-locate software anomalies but also significantly improve the completeness of implicit configuration dependency extraction as well as the quality and efficiency of configuration file generation.

Abstract:To address the issues of information fragmentation and insufficient robustness caused by the modular design of traditional perception systems for unmanned vehicles, this paper proposes a Transformer-based multi-modal BEV fusion algorithm. Firstly, the vision branch, built upon BEVFormer, constructs an image spatio-temporal collaborative encoding mechanism. This is combined with a semantic-guided BEV projection to enhance the representation of key regions, mitigating feature distortion caused by lighting variations and motion blur. Secondly, the radar branch introduces a sparse matrix-state space hybrid model （Sparse-SSM）, which tackles the sparsity and discontinuity of point clouds through sparse voxelization and state transition modeling. Finally, a Transformer cross-attention mechanism is employed to achieve precise alignment and fusion of cross-modal features. The model is trained and experimentally validated on the nuScenes dataset. Results demonstrate that compared to the image-only and point cloud-only branches, the proposed method improves the target detection score by 15% and 23.3%, respectively. Compared to the advanced baseline BEVFusion method, our approach achieves a 2.7% increase in average precision and a 2.3% increase in the target detection score, while reducing floating point operations by 20% through its sparse design and achieving faster inference speed. This provides unmanned vehicles with high-precision, highly robust environmental perception capabilities.

Abstract:A high-input-impedance, low-noise, capacitively-coupled chopper-stabilized instrumentation amplifier for weak bio-signal acquisition is designed in a 180 nm BCD process. Chopper stabilization technique is utilized to mitigate flicker noise and DC offset. To counteract the input impedance degradation introduced by the chopping scheme, a positive feedback loop and pre-charging technique is employed. A DC servo loop is implemented to reject the electrode DC offset. Furthermore, a multi-rate duty-cycled resistor is adopted as the large-value resistor for the integrator within the DC servo loop, aiming to reduce the noise contribution from the DC servo loop itself. Tailored for applications such as electrocardiography or electroencephalography signal processing, the instrumentation amplifier′s -3 dB bandwidth is set to 2 kHz. Post-layout simulation results demonstrate that the amplifier achieves a DC input impedance of 9 GΩ, an input-referred noise of 0.282 μVrms, and a total current consumption of 4.5 μA, corresponding to noise efficiency factor of only 1.63.

Abstract:To improve the speed synchronization performance of dual permanent magnet synchronous motor (D-PMSM) considering parameter mismatches, a speed synchronization control strategy based on ultralocal model-free speed prediction control (UMFSPC) is proposed. Firstly, the mechanical motion equations of the D-PMSM system are established using the ultrarlocal model, so that the use of motor parameters is minimized. Concurrently, a full-order discrete disturbance observer is designed to estimate unmodelled components and disturbance elements within the ultralocal model and a method for controller parameter selection is proposed. Moreover, a cross-coupled structure is employed to compensate currents, enabling coordinated operation of the dual motors. Finally, simulation and experimental comparisons with conventional PI control are carried out. The results show that the proposed strategy effectively improve the synchronization performance caused by motor parameter variations.

Abstract:To address the significant impact of end-face tilt on the measurement accuracy of fiber-optic blade tip clearance sensors, this study presents an enhanced multifiber bundle sensor with improved tilt insensitivity, based on the structure of a dual-ring coaxial fiber-optic sensor. The key innovation involves dividing the six receiving fibers in the inner ring of the dual-ring coaxial probe into two symmetrically distributed receiving fiber bundles. This modification introduces a second independent output characteristic parameter, the ratio of light intensities R2 between the two inner-ring bundles, supplementing the traditional single output parameter R1, which is the ratio of light intensities between the outer and inner rings. Optical simulations with the software Zemax were conducted to verify the relationships between the two enhanced output characteristic parameters (R1, R2), the end-face tilt angle β, and the blade tip clearance z. A dual-parameter decoupling model was subsequently developed to simultaneously invert both the clearance z and tilt angle β based on the measured output parameters. The results demonstrate that when β varies within 0°~10°, the calculated blade tip clearance derived from decoupling the sensor′s output parameters exhibits an error of less than 3% over the measurement range of 1.0~2.5 mm, confirming the reliability and measurement accuracy of the enhanced fiber-optic sensor under complex operating conditions.

Abstract:To address the issues of small and nonlinear output voltage of proton exchange membrane fuel cells, which are susceptible to polarization losses, load fluctuations and insufficient stability when cooperating with Boost converter, this study was conducted. Firstly, a PEMFC output voltage model including three types of polarization losses (activation, ohmic and concentration polarization) and a state-space average model of the Boost converter were constructed. Then, the output function was reconstructed to meet the conditions of feedback linearization, and an inversion sliding mode controller with an adaptive reaching law was designed to suppress chattering. Finally, based on Matlab/Simulink simulation, the performance of BSMC was compared with that of PI control. The results show that the steady-state voltage error rate of BSMC is 0.099% (0.170% for PI control), its dynamic response is 20.1%~34.1% faster, and its voltage drop amplitude and recovery time are superior to those of PI control under disturbance. The results indicates that this strategy can improve the PEMFC voltage control efficiency under complex working conditions and provide theoretical and technical support for its stable operation.

Abstract:The existing deep learning-based channel estimation algorithms have significantly higher accuracy than traditional algorithms, but they are difficult to meet the real-time communication requirements due to their slow operation and large parameters, and cannot be effectively deployed on mobile terminals. To address these issues, a lightweight channel estimation neural network model ESPCN-net is proposed, which models the channel estimation problem as a super-resolution task of reconstructing a high-resolution image (i.e., the complete channel response) from a low-resolution image (i.e., the channel response at sparse subcarriers). This model abandons the interpolation step in existing neural network models and directly learns the spatial correlation of channel features from pilot signals without interpolation preprocessing to restore high-precision channel state information. Experimental results show that in typical 5G OFDM multipath channel scenarios, compared with SRCNN and ChannelNet, the proposed model improves the computational efficiency by 25.7% and 77.2 times, respectively. In terms of computational complexity, the number of multiply-accumulate operations is only 36.4% of SRCNN and 0.76% of ChannelNet, demonstrating excellent computational efficiency. In terms of parameter scale, although the 47.25×103 parameters of ESPCN-net are more than those of SRCNN, they are far less than those of ChannelNet (682.34×103) and ReCNet (100.45×103).

Abstract:In order to enhance the reliability and efficiency of LDPC code decoding in MLC NAND flash memory, this paper proposes the MSWBF decoding algorithm. This algorithm no longer relies solely on the minimum LLR value as the basis for flip decisions; instead, it performs linear integration of the minimum LLR value associated with check nodes and variable nodes with their average LLR value, thus constructing a more stable local confidence estimation model. Furthermore, an exponential adjustment function is introduced in the calculation of flip indicators for variable nodes to dynamically strengthen the enhance the identification capability for low-confidence bits. Combined with the physical characteristics of flash memory, this paper establishes a channel model that includes RTN, DRN, CCI and creates generates soft information inputs through Monte Carlo simulation. Based on this, QC-LDPC codes are constructed and a decoding architecture is designed, comparing the performance of algorithms such as BF, WBF, MWBF, IMWBF and BP. Simulation results show that MSWBF improves the decoding success rate by 556% under strong interference conditions, and the average number of iterations decreases by over 30%, effectively improving decoding robustness and convergence efficiency. This algorithm is hardware-friendly and exhibits strong potential for practical implementation, providing a feasible improvement path for low-complexity soft decision decoding methods.

Abstract:Steady-state visual evoked potentials (SSVEP) have become one of the most widely used EEG signals in brain-controlled robot systems due to their advantages of high signal-to-noise ratio （SNR） and high information transmission rate. However, due to the phenomenon of ′BCI blindness′, some subjects have difficulty inducing stable SSVEP responses, which limits the application of this technology in practical scenarios. To improve the recognizability of SSVEP signals, this paper proposes a novel modulation SSVEP paradigm that integrates event-related potential time series, and designs an online brain-controlled robot system based on this paradigm, which can achieve coordinated control of wheeled robots in complex scenarios. Firstly, without changing the standard SSVEP frequency characteristics, embedding event-related potential stimuli can enhance user attention and improve signal quality; then, EEG signals generated by 8 subjects under the modulation paradigm and standard paradigm were collected. By comparing their offline analysis results under the filter group canonical correlation analysis and task correlation analysis algorithms, the improvement effect of modulation paradigm on EEG signals was verified, and the algorithm with high accuracy was selected for data processing in the online system; finally, this paper develops an online brain-controlled robot system based on this paradigm and conducts online experiments. The offline analysis results show that the modulation paradigm generates EEG signals with a classification accuracy improvement of 2%~32% compared to the standard SSVEP paradigm under various classification algorithms, effectively alleviating the problem of ′BCI blindness′. The online brain-controlled robot system has shown good classification performance in multiple subjects, with an average information transmission rate （ITR） of 30.74 bits/min, proving its feasibility and effectiveness in practical applications.

Abstract:To address the characteristics of complex backgrounds and multi-scale objects in remote sensing images, a remote sensing image object detection algorithm based on YOLOv8n, named PAM-YOLO, is proposed. First, in the backbone network, a triple pool hybrid attention (TPHA) mechanism is constructed and embedded. By fusing information from global maximum, average, and median pooling, it effectively suppresses the interference of background noise on channel features and enhances the feature representation of key regions. Second, a spatial context aware module (SCAM) is introduced in the neck structure. By mining global contextual information and building long-range dependencies between channel semantics and spatial structure, it improves the model′s ability to distinguish between targets and backgrounds. Finally, a parallel branch feature extraction (PBFE) module is designed and introduced. Through the deep decoupling and interactive fusion of three parallel branches for local details, global semantics and multi-scale context, it achieves a refined representation for objects of different scales. On the DIOR dataset, PAM-YOLO′s mAP50 increased by 1.8%, mAP50.95 by 2.4%, P-value by 0.3%, and R-value by 1.8%; its metrics on the DOTA dataset are also superior to other algorithms. The experimental results indicate that PAM-YOLO shows higher detection accuracy and robustness in the task of remote sensing image object detection.

Abstract:To address the challenge of missed detections for distant small objects and occluded targets in complex autonomous driving scenarios, this paper proposes an improved algorithm based on YOLO11n. First, an efficient multi-scale attention mechanism is introduced to replace the original C2PSA module in the backbone network. It performs group reshaping and parallel multi-scale fusion without reducing channel dimensions, thereby enhancing the model′s focus on small objects. Second, an additional 160×160 detection layer is incorporated to leverage detailed spatial information from shallow features for the precise localization of small targets. Furthermore, a context-guided module, designated C3k2-CGB, is designed to augment feature fusion with global contextual information, improving the recognition capability for occluded objects. Finally, the Wise-IoU loss is adopted for bounding box regression, which suppresses the harmful gradients from low-quality examples. Experimental results on the KITTI dataset demonstrate that the improved model achieves a 5.9% increase in mAP and a 10.3% gain in recall, while reducing the number of parameters by 10.1%. It outperforms several mainstream YOLO variants, showing significant improvements in detecting small and occluded objects.

Abstract:Aiming at the problem of degraded measurement accuracy in air data systems （ADS） caused by airflow separation and shock wave interference during complex flight conditions such as transonic regimes and high maneuverability, this paper proposes an atmospheric parameter reconstruction method based on datadriven wind field estimation combined with velocity vector triangle resolution. Utilizing historical flight data, the method employs a least squares polynomial fitting algorithm for wind speed prediction and leverages a velocity vector triangle model to inversely compute true airspeed, thereby reconstructing key atmospheric parameters including angle of attack (AOA), angle of sideslip (AOS), and Mach number. Simulation results demonstrate that the proposed method achieves significant accuracy improvement in atmospheric parameter reconstruction under complex conditions: The original AOA measurements fluctuate around 2°, while the original AOS measurements fluctuate around 1°. Evaluated using mean absolute error (MAE) and root mean squared error (RMSE) metrics, the errors are 0.47° (MAE) and 0.50° (RMSE) for AOA, 0.20° (MAE) and 0.24° (RMSE) for AOS， 0.005 5 (MAE) and 0.006 3 (RMSE) for Mach number. Additionally, the study compares the proposed method with an atmospheric parameter estimation approach based on the frozen wind field theory, with both methods effectively reducing measurement errors in AOA and AOS. This research provides a software solution for high-precision, highly reliable air data measurement in complex flight environments without requiring additional hardware.

Abstract:In response to environmental pollution and the energy crisis, countries worldwide are actively advancing green wind power. However, owing to uncontrollable variables such as wind speed and direction, wind power generation is characterized by volatility and intermittency, necessitating accurate forecasting of wind power output to ensure the secure operation of integrated energy systems. To mitigate this challenge, this paper proposes a wind power forecasting method based on a dual-tower Transformer model, which integrates multiple features. This approach thoroughly investigates wind power data enriched with meteorological information, extracting feature information comprehensively. The method incorporates temporal attention, empirical distribution, and K-means clustering to construct a Transformer model that decouples meteorological patterns by considering temporal dependencies and distributional traits. Additionally, feature-level attention, Hilbert transform, and seasonal trend decomposition are integrated to build a Transformer model for seasonal trend decomposition that leverages feature associations and frequency-domain characteristics. Finally, a weighted average fusion module is proposed to combine the predictions of the two aforementioned Transformer models, thereby enhancing the accuracy of wind power forecasting. Extensive experimental comparisons with advanced algorithms from the past five years, conducted on both general public datasets and proprietary datasets from enterprises, reveal that the proposed method markedly outperforms existing techniques in terms of both prediction accuracy and model interpretability. In comparison with the runner-up algorithms, the proposed method reduces the mean squared error (MSE) and the mean absolute error (MAE) by up to 15.34% and 6.04%, respectively, while enhancing the coefficient of determination R2 by as much as 39.33%.

Abstract:Aiming at the problems that the standard hippopotamus optimization algorithm (HO) has slow convergence speed, easy to fall into local optimum and imbalance between search and development, this study proposes the multi-strategy cooperative hippopotamus optimization algorithm (MSCHO). Circle mapping is introduced in the initialization stage to enhance the diversity of the initialization population and reduce the blind area of initialization; secondly, the mirror-pole random wandering strategy is introduced in the HO position updating stage to enhance the possibility of finding the optimal solution; finally, nonlinear adaptive Cosine adaptive t-distribution variation is used in the HO exploration stage to improve the convergence speed and stability of the algorithm. In order to verify the improvement effect and performance of the algorithm, five swarm intelligent optimization algorithms are selected to simulate eight classical test functions and the experimental results show that compared with other algorithms, the improved MSCHO has more excellent optimization ability and convergence speed.

Abstract:A method using a calibration device consisting of two orthogonal planes was proposed to address the moving direction calibration problem for 2D LiDAR mounted on a linearly moving platform. When the linearly moving 2D LiDAR performs two or more scans of the calibration device, geometric constraints on the motion parameters can be derived from the scan lines on the two calibration planes. The parameters must lie on a hyperbola on the parameter plane. Additional constraint hyperbolas are generated by adjusting the pose of the calibration device and conducting new scans. To solve the calibration problem, four algorithms were presented: Linear least squares method (LSM), Gauss-Newton method (G-N) and an optimization approach that minimizes the sum of squared shortest distances from a point to all constraint curves (denoted as Opt-D), as well as a method that optimizes the orthogonality of the reconstructed calibration planes (denoted as Opt-O). Simulation experiments demonstrated that the Opt-O method achieves the highest accuracy. In real experiments, a 3D point cloud of a plastic ball was constructed from the calibration results and was used to estimate the spherical equation. When using the results of G-N, Opt-D, or Opt-O methods, the mean distances between the reconstructed sampled points and the fitted sphere are all less than 0.5 mm, with the standard deviations all less than 0.5 mm. These results collectively validate the effectiveness of the proposed calibration method.

Abstract:To improve the accuracy of on-orbit object detection in remote-sensing images under complex scenarios, this paper proposes a detection method based on multi-scale feature fusion. First, a multi-scale feature fusion module is designed to expand the model′s receptive field, where parallel multi-channel fusion is employed to extract multidimensional target features. Second, a spatial-channel feature interaction module is constructed by integrating an attention mechanism to enhance the model′s ability to capture detailed features in salient regions. Third, the normalized Wasserstein distance is introduced to improve the loss function, optimizing the measurement of positional offsets between overlapping anchor boxes and enhancing discrimination of densely overlapped targets. In addition, a ship dataset named JS25k is annotated on “Jilin-1” satellite imagery with data augmentation and other preprocessing steps. The proposed model achieves a final detection accuracy of 98.2%, outperforming the latest YOLOv13 by 0.6% and exhibiting better performance than the RT-DETR series. When deployed on an edge-side embedded platform, the detection speed reaches 92.3 fps. Experimental results demonstrate that the model effectively improves detection accuracy in complex scenes while meeting real-time requirements for on-orbit applications.

Abstract:Advances in high spatiotemporal resolution millimeter-wave radar technology have significantly improved range and velocity resolution, thereby causing targets to span multiple adjacent cells in the range-Doppler map and form extended structures. These structures violate the point-target assumption underlying conventional constant false alarm rate algorithms, leading to inaccurate background noise estimation and consequently, incomplete detection of extended targets. To address this issue, this paper proposes a millimeter-wave radar detection method based on convolution and entropy-guided weighted sampling. The method first performs region enhancement on the range-Doppler map using a cross-shaped convolutional kernel, which reinforces spatial continuity within target regions. It then estimates the global background noise using weighted random sampling guided by the entropy distribution across the range-Doppler map, which replaces the conventional sliding reference window structure and achieves more accurate background noise estimation. Simulation results demonstrate that the proposed method achieves 90% detection probability at an signal-to-noiseratio(SNR) of -5 dB, showing a notable improvement in extended target detection performance over existing methods. Field experiments further verify its effectiveness in preserving target structural integrity and enhancing point cloud density. In two representative traffic scenarios, the proposed method increases the number of detected points by 21.51% and 111.11%, and 39.38% and 136.87%, respectively, indicating its potential for robust millimeter-wave radar perception in complex environments.

Abstract:In the magnetic flux leakage (MFL) detection process of oil pipelines, the received MFL signals have a relatively low signal-to-noise ratio (SNR) and are often aliased with other low-frequency noises that have similar frequency components. Currently, common denoising algorithms cannot effectively separate MFL signals from other noise components. To improve the accuracy of MFL signal extraction, this paper proposes a method combining variational mode decomposition (VMD) optimized by the giraffe optimization algorithm (GOA) and dynamic wavelet threshold denoising (DWTD) for MFL signal identification.Firstly, the GOA algorithm is used to select the input parameters of VMD. Then, VMD performs adaptive decomposition on the signal based on the optimized parameters to obtain a certain number of modal components, which are divided into effective signal components and noise components using correlation coefficients.Finally, dynamic wavelet threshold denoising is applied to the effective signal components to obtain the noise-removed MFL signals.Tests and analyses were conducted on simulated signals with different intensities and actual measured MFL signals. The index analysis of simulated signals shows that the SNR of the proposed method is at least 20% higher than that of other methods, and the smoothness of actual measured MFL signals is 10% higher than that of other methods. The results indicate that the denoising method combining GOA-optimized VMD parameters and DWTD can effectively remove interference noise and retain the original signal waveform more completely. This method outperforms other commonly used denoising methods and is suitable for denoising MFL signals of oil well pipelines.

Abstract:In response to the shortcomings of traditional fault detection methods in terms of deployment difficulty, environmental anti-interference capability, and early fault sensitivity, this paper proposes a voiceprint fault detection model based on WSSA-VMD and the improved Swin Transformer. First, a sparrow search algorithm integrated with the whale optimization algorithm (WSSA) is introduced to adaptively optimize the key parameters of variational mode decomposition (VMD), thereby enhancing the denoising performance of the original acoustic signals. Next, the effective intrinsic mode functions obtained from the optimized VMD are selected to reconstruct the signal, which is then converted into Log-Mel spectrograms to serve as high-quality feature inputs. Finally, an improved Swin Transformer model is constructed by incorporating the convolutional block attention module (CBAM) and replacing the original normalization layer with a root mean square normalization method, so as to strengthen the model′s feature extraction capability and training stability. Experimental results demonstrate that the proposed model achieves an overall accuracy of 97.93%, effectively identifying incipient blade faults while exhibiting strong generalization performance.

Abstract:Traditional wood classification methods face significant challenges, including low efficiency, heavy reliance on manual expertise and prevalent material fraud in furniture markets. Additionally, there exists a research gap in the application of X-ray single-projection imaging for wood identification. To address these issues, a set of special imaging device for wood material visualization was independently developed and an intelligent wood authentication system that integrates X-ray single-projection imaging with a dual-attention mechanism was proposed. A systematic X-ray single-projection dataset comprising 10 typical commercial wood species was constructed, including three Pterocarpus species, four Dalbergia species, and three other woods. A key innovation of this study is the proposed EDMA-Net model. This architecture integrates a hierarchical GAM-ECA dual-attention mechanism that enables synergistic optimization of both shallow global texture features and deep channel characteristics. Experimental results indicate that the model achieves an overall classification accuracy of 98.04%. Ablation studies show that, compared to the baseline model, EDMA-Net boosts the classification accuracy by 2.38% while compressing the parameters to only 16.51 M. The average single-image inference time is only 7.00 ms, reaching an inference speed of 142.95 fps, thereby verifying the module′s effectiveness. Furthermore, actual furniture panels tests were conducted and the accuracy for all samples exceeded 95%. By leveraging X-ray single-projection imaging technology, this system provides a non-destructive, rapid and practical technical solution for material authentication, quality monitoring, and wood processing management in the furniture market.

Abstract:To address the issues of weakened defect features and localization difficulties in wind turbine blade drone inspections, this paper proposes a lightweight, highprecision detection algorithm based on the improved YOLO11 framework. Firstly, a mixed local channel attention (MLCA) module is introduced into the backbone to collaboratively model local texture details and global context, enhancing sensitivity to minor defects under complex lighting. Secondly, ADown progressive downsampling replaces standard convolutions to effectively preserve defect edge integrity while reducing feature map resolution. Finally, a multi-scale attention feature pyramid network (MAFPN) is constructed to improve semantic representation through cross-level feature aggregation. Experimental results on a real-world inspection dataset demonstrate that the proposed method achieves mAP@0.5 and mAP@0.5:0.95 scores of 85.7% and 61.1% (improvements of 4.0% and 5.6%, respectively), while reducing parameters by 22.5% and GFLOPs by 9.5%. The conclusion indicates that the algorithm significantly reduces missed and false detections in complex scenarios. Compared with representative models such as I-YOLOv8n and O-YOLO11, it achieves a superior accuracy-efficiency trade-off, making it highly suitable for deployment on edge devices like UAVs.

Abstract:To address the challenges of low detection accuracy for small-scale defects and excessive model parameters in wind turbine blade surface inspection, this paper proposes an improved detection framework, CESL-YOLO, based on YOLOv8n. First, the convolutionalblockattention module（CBAM） is integrated into the backbone to enhance the model′s focus on defect regions during feature extraction. Second, the neck structure is redesigned as an EfficientFPN, where re-parameterized convolution is employed to improve multi-scale feature fusion efficiency. Third, the SIoU bounding box regression loss function is adopted to accelerate model convergence. Finally, a lightweight sparse-separable branch (LSSB) detection head is developed to improve defect localization accuracy. Experimental results demonstrate that CESL-YOLO outperforms the baseline model, achieving gains of 2.7%, 4.8% and 5.3% in precision, recall and mAP, respectively. Moreover, the model size is reduced from 3.01 M to 2.82 M parameters, making it more suitable for practical wind turbine blade surface defect detection tasks.

Abstract:The capacity to detect craters in complex scenes is of paramount importance for the fields of deep space exploration and planetary science. In order to address the accuracy bottleneck in detecting degraded craters within rille regions and at crater margins, this paper proposes the HCF-YOLO crater detection network, which synergistically integrates discrete wavelet transform and cross-attention. The wavelet-domain preserved downsampling module is designed to mitigate the decay of structural features in degraded crater edges. This enhancement of distinguishability from the background is accompanied by a reduction in false negative rates. Secondly, a guided cross-attention fusion module is introduced. The generation of weights from shallow-level crater structural details is achieved in order to directionally constrain cross-scale fusion with high-level semantics. The suppression of rille texture noise interference is of particular significance in this context, as it prevents clear, large craters′ semantic features from masking structural features. This, in turn, serves to reduce false positives caused by background noise and feature confusion. The two modules have been shown to enhance detection robustness in complex scenes in a synergistic manner. Finally, systematic comparisons and ablation experiments are conducted on the MDCD dataset. In comparison with state-of-the-art methods, true positive (TP) detections increase by 133, recall improves by 5%, and precision rises by 2%.

Abstract:To address the issues of large parameter counts, insufficient detection accuracy, and difficulty in deployment on embedded devices in existing strip steel surface defect detection algorithms, a lightweight detection algorithm named YOLO-AMM based on improved YOLOv11n is proposed. Firstly, the ADown module is introduced to solve the problem of tiny defect feature loss caused by traditional strided convolution downsampling, thereby improving accuracy and to reduce the number of parameters for enhanced real-time performance. Secondly, the C3k2_MDSB module is constructed to strengthen multi-scale feature extraction capability and computational efficiency. Finally, the mixed local channel attention （MLCA） attention mechanism is incorporated into the feature fusion network to improve the feature extraction ability for defects of different scales. Experimental results show that compared with YOLOv11n, the mAP50 of the YOLO-AMM algorithm is increased by 3.6%, the recall rate by 2.3%, and the precision by 2.9%. Meanwhile, the number of parameters is reduced by 18.0%, the computational load by 14.1%, and the model size by 16.5%. The proposed algorithm outperforms a variety of classic lightweight algorithms and similar improved algorithms and is suitable for deployment on resource-constrained embedded devices.