Abstract:With the increasing penetration of renewable energy and power electronic devices in power systems, this paper proposes a virtual synchronous generator frequency regulation strategy that combines a radial basis function adaptive method with model predictive control, using a photovoltaic-energy storage system as an example. Based on the dynamic adjustment of system rotational inertia and damping coefficient, a three-step frequency prediction model is established, introducing the system angular frequency deviation and active power as external inputs. Through rolling optimization by the model predictive controller, the compensation power of the virtual synchronous generator is optimized to dynamically correct the reference power. On the DC side, the photovoltaic and energy storage units jointly supply power, ensuring stable active power output of the virtual synchronous generator and maintaining DC bus voltage regulation. Simulation results demonstrate that the proposed strategy effectively suppresses active power oscillations in the photovoltaic-energy storage grid-connected system, reduces frequency deviation and its rate of change, and meets the frequency regulation requirements of high-penetration renewable energy systems.

Abstract:Short-term power load forecasting faces challenges such as multiple influencing factors, nonlinear and non-stationary load data, which make traditional prediction methods insufficient for engineering applications. To address this, this paper proposes a combined prediction model based on MPSR and IHBA-BiGRU-Attention. First, Spearman correlation coefficients are used to select power load influencing factors, forming a multivariate time series with load sequences; subsequently, multivariate phase space reconstruction (MPSR) is applied to this time series as input variables. An improved honey badger algorithm (IHBA) integrating Kent chaotic mapping, density factor modification and small-hole imaging opposite learning strategy is employed to optimize hyperparameters of the bidirectional gated recurrent unit (BiGRU) network, while the attention mechanism is applied to dynamically weight hidden state sequences of BiGRU, constructing a short-term power load prediction model based on MPSR and IHBA-BiGRU-Attention; finally, comparative experiments on power load datasets from Australia and Singapore showed mean absolute percentage errors of 0.650% and 1.067%, respectively, both lower than other models, validating its effectiveness. Additionally, predictive experiments on actual load datasets from a southern region in China achieved a mean absolute percentage error of 1.523%, outperforming competing models and demonstrating the engineering applicability of the proposed model.

Abstract:In precision measurement tasks such as augmented reality (AR) displays, laser beam calibration, and diffractive optical element inspection, the sub-pixel localization accuracy of spot centers is critical for system stability and reliability. Conventional positioning methods based on model fitting or rule extraction struggle to achieve robust and accurate spot center estimation in complex scenarios due to system aberrations, diffraction distortions, and structural noise. To address these limitations, a U-Net-based spot center localization framework is proposed. This framework incorporates an adaptive dual-channel heatmap supervision mechanism to simultaneously model the spot center and its diffusion structure, enabling precise characterization of real-world point spread function (PSF) features. During inference, a Hessian matrix-based second-order differential extremum detection method is employed to enhance the stability of sub-pixel peak localization. Experimental results demonstrate that the proposed method exhibits a concentrated error distribution on real optical datasets, achieving an RMSE of 0.413 pixel. Compared with HRNet and FFD, the proposed method significantly improves both localization accuracy and stability.

Abstract:UAV-assisted communication has broad application prospects in future wireless communications. Deploying it in wireless powered communication networks enables energy transmission and information collection for nodes. This paper introduces a magnetic coupling resonant wireless power transfer method based on the parity-time symmetry principle, and adopts a dual-UAV cooperative energy supply communication system. With the goal of maximizing the minimum sum rate, the trajectory of the information-collecting UAV and the power allocation of nodes are jointly optimized. Since this problem is non-convex, it is solved by methods such as the block coordinate descent algorithm, weighted minimum mean square error algorithm, and successive convex optimization. The power-transfer UAV uses a dynamic programming algorithm to select the shortest path traversing all nodes. Simulation results show that the proposed joint optimization algorithm has good convergence. Compared with the trajectory planning scheme, the minimum sum rate is improved by more than 2 times; compared with the straight-line flight scheme, the minimum sum rate is increased by about 10.7% and 4.2% under different cycles, respectively; compared with the radio frequency power transfer scheme, the minimum sum rate is improved by more than 7 times. These results verify the applicability of the proposed scheme in UAV-assisted wireless energy supply communication networks.

Abstract:To address the non-stationary characteristics and multi-scale dynamic evolution of wind power time series, this study proposes a hybrid prediction model named VMD-FECAM-TCN-NTSformer. Firstly, Variational Mode Decomposition （VMD） is employed to adaptively decompose the original power sequence in the frequency domain, effectively separating noise interference and extracting multi-scale dynamic intrinsic mode functions. Secondly, Temporal Convolutional Networks （TCN） are used to hierarchically capture local temporal features through dilated convolutions. Meanwhile, the NTSformer utilizes a de-stationary attention mechanism to dynamically correct normalization bias, thereby enhancing the modeling of abrupt trends and periodic fluctuations. Furthermore, a Frequency Enhanced Channel Attention Module （FECAM） is introduced to extract frequency-domain features via fast Fourier transform and dynamically assign channel weights to focus on key frequency components. Experimental results show that the proposed model improves the coefficient of determination R2 by more than 17% compared to traditional CNN models and by more than 5% compared to Transformer-based models, in both 15-minute single-step forecasting and 60-step (15 h) multi-step forecasting. These results demonstrate the model′s superior prediction accuracy and robust forecasting performance.

Abstract:UAV infrared image small target detection is a key technology in the fields of reconnaissance and surveillance, search and rescue, and its reliability is challenged by problems such as small target size, weak feature expression, and complex background. To solve the above problems, this paper proposes a UAV infrared small target detection algorithm PGF-RTDETR based on multi-scale perception and adaptive screening strategy. Firstly, a polarized channel-gated unit is constructed in the neck network, and the polarized attention mechanism is used to guide multi-scale features for deep semantic fusion, so as to enhance the recognizability of small targets and effectively suppress background interference. Secondly, an adaptive convolution enhancement module is designed to enhance the selectivity and fineness of features by using the gating mechanism to improve the extraction efficiency of fine-grained features. Finally, a high-efficiency residual structure that integrates partial convolution and channel transformation is introduced into the backbone network, which improves the feature expression ability and effectively reduces the number of model parameters. Experimental results show that compared with the benchmark model RT-DETR-R18, PGF-RTDETR increases mAP50 and P by 2.6% and 6.6% on the HIT-UAV dataset, and reduces the number of model parameters and computation by 20.7% and 25%, respectively, and mAP50 increases by 0.6% on the FLIR dataset. The improved model improves the detection accuracy while maintaining a small amount of parameters and computational amount, and provides an effective solution for UAV infrared small target detection.

Abstract:In order to guide electric vehicle users to participate in orderly charging, this paper proposes an internal evolutionary game model of the electric vehicle group based on multi-strategy fusion. Firstly, an evolutionary game model including three strategies: Charge-only, charge-discharge, and peak-time-avoiding charging was designed. A dynamic threshold mechanism was introduced, combined with the empirical weighted attraction rule and the Fermi rule, to achieve multi-strategy fusion and update. By setting discharge incentives and charging subsidies, users′ enthusiasm for choosing charging and discharging strategies has been enhanced. Secondly, this paper establishes a game model within the electric vehicle group, analyzes the benefits of users under different strategies, and verifies the validity of the model through simulation. Finally, the multi-strategy fusion update rule method was adopted. Through the dynamic threshold mechanism, combined with the long-term learning ability of EWA and the shortterm response advantage of Fermi, the stability and adaptability of strategy adjustment were significantly improved. The simulation results show that this model can accurately simulate the decision-making behavior of users, effectively improve the efficiency of power distribution, and reduce the peak-valley difference of the power grid by approximately 36.7%. Compared with the update rule of a single strategy, the multi-strategy fusion method shows significant advantages in the stability and adaptability of strategy adjustment. This article provides theoretical support and reference methods for power grid operators to formulate incentive measures and for electric vehicles to participate in vehicle-to-grid interaction.

Abstract:Traffic cones are commonly used to mark temporary roads in emergencies like traffic accidents or road control. However, current autonomous driving research focuses mainly on structured roads, with relatively little on unstructured roads. Thus, enabling unmanned vehicles to accurately perceive temporary roads formed by traffic cones in such environments is vital for improving driving safety. This study proposes a deep learning model based on multi-sensor data fusion to detect temporary roads by quickly identifying traffic cones in special scenarios. The framework uses an improved YOLOX model to obtain traffic cones color information via machine vision, then fuses this with the distance data of traffic cones detected by LiDAR to achieve real-time perception of temporary roads. Experimental results show that in unstructured road environments, the model has a 31 ms single-frame detection time and over 85% accuracy within 15 meters. It enables real-time traffic cone detection, meets design goals, and is of great significance for enhancing the driving safety of unmanned vehicles.

Abstract:To achieve more efficient and precise control of permanent magnet synchronous motors (PMSM) in the field of electronic measurement, an active disturbance rejection controller (ADRC) based on the multi-strategy improved black kite algorithm (MBKA) was proposed to address the challenges of PMSM, such as numerous nonlinear active disturbance rejection control parameters, difficult tuning, and chattering issues in traditional nonlinear functions. First, the original black kite algorithm (BKA) was enhanced by initializing the population with chaotic mapping and improving the position update process using the cross-over and mutate strategy and adaptive Cauchy-Gaussian walk. The proposed improvements demonstrated optimal performance on test functions. Subsequently, the multi-strategy improved black kite algorithm was applied to optimize the ADRC. After conducting a Sobol parameter sensitivity analysis on the ADRC, the parameters to be optimized were selected. The improved black kite algorithm achieved faster convergence of ADRC parameters and stronger capability to escape local optima. When the improved strategy was applied to PMSM motor field-oriented control (FOC), the motor current harmonics were reduced, system efficiency improved, dynamic response accelerated, and disturbance rejection performance enhanced.

Abstract:To address the challenges of complex background, dense small target distribution, and diverse target scales in remote sensing images, an improved algorithm based on YOLO11 was proposed. First, ghost convolution was introduced into the backbone network to significantly reduce computational load and parameter count while maintaining model performance. Second, a hybrid network module was designed in the backbone, incorporating three types of modules to enrich information flow and enhance feature extraction capabilities. Finally, a semantic-aware cross-layer multi-feature fusion approach was adopted, replacing the original P5 layer with the P2 layer to strengthen multi-scale feature fusion capabilities, effectively improving detection accuracy and mitigating the difficulty in extracting small target features from remote sensing images. Experimental results on the VisDrone2019, AI-TOD and RSOD datasets demonstrated that the improved YOLO11s model achieved mAP50 increases of 2.8%, 5.1% and 5.3% respectively compared to the original YOLO11s, with a 31% reduction in parameters, validating the effectiveness of the new algorithm.

Abstract:In outdoor parking lots, due to the uncertainty and dynamic interaction of vehicle spatiotemporal distribution and the limitations of real-time computing resources, the real-time performance and efficiency of omnidirectional mobile robots in dynamic parking lot path planning need improvement. Current path planning algorithms struggle to meet the demands of such real-time dynamic environments. This paper proposes an improved A*-DWA algorithm incorporating fuzzy control, which combines fuzzy global planning driven by environmental obstacle factors with local dynamic trajectory optimization using the dynamic window approach to address the path planning challenges of omnidirectional mobile robots in real-time dynamic environments. First, an expanded 20-neighborhood search strategy is adopted to optimize global path generation. Second, a fuzzy adaptive evaluation mechanism based on multi-parameter weighting is designed to enhance the algorithm′s adaptability to dynamic environments. Then, cubic spline interpolation is employed to achieve path smoothing. Finally, DWA is integrated for local dynamic obstacle avoidance and trajectory optimization. By constructing a simulated outdoor parking lot environment, the results demonstrate that compared to the algorithms proposed in the literature, the FCA*-DWA algorithm improves path length, search efficiency, and node optimization by 13.30%, 21.16% and 45.45%, respectively, providing methodological guidance for autonomous navigation of mobile robots in complex dynamic scenarios.

Abstract:To address challenges including rule optimization difficulty, complex parameter tuning, and poor multi-scenario adaptability in existing fuzzy logic evaluation methods for complex industrial applications, this paper proposes an intelligent assessment system based on double-optimized fuzzy inference. First, a double optimization fuzzy inference system is designed, integrating fuzzy entropy-guided rule generation with fuzzy gradient collaborative descent to alternately optimize model structure and parameters, coordinated dynamically by an adaptive optimization scheduler. Second, a multi-scale fuzzy feature extraction network and adaptive fuzzy feature fusion mechanism are constructed, where the former extracts multi-granularity fuzzy features through parallel multi-scale branches while the latter achieves intelligent feature fusion via fuzzy channel-spatial co-attention. Finally, a dynamic fuzzy weighting allocation is proposed, employing a scene-aware weight generation network to dynamically adjust fuzzy rule weights based on input features. Validated in natural gas pipeline risk assessment and electrical equipment identification scenarios, experimental results demonstrate 95.83% assessment accuracy,for pipelines, and 96.54% accuracy with 96.32% F1-score for equipment identification. Compared to conventional fuzzy logic and deep learning methods, the proposed approach significantly enhances evaluation accuracy and generalization capabilities.

Abstract:In UAV traffic surveillance, small target detection faces challenges such as insufficient feature representation and low efficiency in multi-scale fusion. To address this, this study proposes YOLO-MAF, a lightweight detector. First, the multi-scale edge enhancement (MSEE) module strengthens edge information via an adaptive multi-scale receptive field. Second, the SEGE module combines soft nearest-neighbor interpolation (SNI) and enhanced group convolution (GSConvE) to improve cross-level alignment and fusion. Finally, the MASF-Head adopts dual attention to learn spatial-channel weights for adaptive multi-scale fusion. On VisDrone2019, YOLO-MAF achieves 45.6% mAP@0.5 and 29.4% mAP@0.5:0.95, improving the baseline by 7.3%and 6.4% with 50% fewer parameters, demonstrating effective small-object detection under UAV scenarios.

Abstract:For the task offloading problem of minimizing latency in mobile edge computing, a quantum genetic algorithm is proposed to achieve the minimum total latency of computing tasks under energy consumption constraints. Firstly, with the offloading strategy minimizing the total latency as the optimization objective, a system model and a computing model are established; secondly, by balancing the conditions of minimizing total latency with energy consumption and computing resources constraints, a target function for minimizing total latency is constructed; finally, by combining the quantum genetic algorithm with quantum mutation gates and elite retention strategies, an offloading strategy for minimizing total latency is obtained. Simulation experiments show that compared with other offloading strategies, the proposed offloading strategy has a smaller total latency under different energy consumption constraints and different numbers of computing tasks, and its total latency is reduced by 7.3% compared with the genetic algorithm and by 4.3% compared with the adaptive particle swarm optimization algorithm under different task data volumes.

Abstract:Direction of arrival （DOA） estimation is an important research area in array signal processing, where estimation accuracy is closely related to the array aperture. Increasing the array aperture can effectively improve DOA estimation performance. However, traditional methods usually rely on increasing the number of array elements to expand the aperture, which is limited by physical size and hardware costs in practical applications. Therefore, effective expansion of array aperture without increasing physical resources is worth studying. This paper proposes a multi-feature fusion based learning method for array aperture expansion. By employing a multi-scale convolution module to extract features from the received signals of a small-aperture array, and combining it with a channel attention module for adaptive weighted fusion of multiple features, the proposed method ultimately generates received signals of a larger-aperture array. Simulation results show that the proposed method can expand the array aperture based on single snapshot signal of small aperture array and significantly improves DOA estimation performance.

Abstract:To address the limitations of traditional fault diagnosis models in collaboratively extracting spatiotemporal features from gearbox vibration signals and their weak robustness in noisy environments, this paper introduces an intelligent fault diagnosis method that combines a receptive field attention residual network (ResNet-RFA) and a bidirectional gated recurrent unit with self-attention (BiGRU-SATT). The process begins by converting raw vibration signals into time-frequency images via short-time fourier transform (STFT), while preserving the original 1D time-series data. A dual-channel network is then constructed: One channel uses ResNet-RFA to extract key spatial features from the time-frequency images, and the other uses BiGRU-SATT to capture temporal dependencies. The spatiotemporal features are merged and fed into a fully connected layer for classification. Experimental results demonstrate a high accuracy of 100%, outperforming comparison models (Transformer: 91%, Mamba: 96%, SVM: 94%, DBN: 89%) and showing strong noise robustness under 10 and 20 dB Gaussian-impulse mixed noise. In conclusion, the fusion model of ResNet-RFA and BiGRU-SATT can effectively and collaboratively mine the spatiotemporal features of signals, demonstrating superior accuracy and robustness over other comparative models, making it suitable for complex industrial environments.

Abstract:Cellular vehicle-to-everything(C-V2X), as a representative communication technology for intelligent connected vehicles, enables data interaction between vehicle-to-vehicle, vehicle-to-infrastructure, and vehicle-to-pedestrian by establishing highly reliable communication links. However, the complex and variable communication environment of the internet of vehicles poses significant challenges to terminal identity access. Due to its characteristics of uniqueness, stability, and unclonability, radio frequency fingerprint (RFF) can provide a physical layer security solution for C-V2X terminal identity access. Based on this, this paper proposes a C-V2X terminal radio frequency fingerprint extraction and authentication scheme: First, an effective preprocessing algorithm is designed to separate the demodulation reference signal (DMRS) in the physical sidelink shared channel (PSSCH) and the physical sidelink control channel (PSCCH); a logarithmic spectrum separation algorithm is adopted to suppress the noise components caused by the randomization of the DMRS sequence; a training method based on the siamese adversarial network (SANet) is designed to make the feature extraction network focus on extracting hardware-related device fingerprints. Experimental results show that the designed preprocessing and logarithmic spectrum denoising algorithms can effectively improve the stability and recognition accuracy of terminal fingerprints in multiple scenarios; the SANet exhibits excellent generalization ability in cross-channel environment tests: The average authentication precision and recall reach 93.22% and 92.67% in static scenarios, and 82.85% and 82.21% in mobile scenarios, respectively.

Abstract:Aiming at the supply chain security risks and technical bottlenecks caused by the import of key components of domestic instrument landing system （ILS）, this paper proposes a design scheme of ILS signal transmitting unit based on all-domestic chips. The design adopts domestic FPGA and DDS technology to realize the high-precision generation of 90, 150 Hz single tone signals and 1 020 Hz Morse code identification signals. Digital-to-analog conversion of baseband signal is completed by domestic DAC, and CSB/SBO signal synthesis is realized by I/Q modulation architecture combined with RF carrier generated by frequency source devices. The power dynamic regulation is completed by combining digital attenuator and high linear power amplifier. Finally, according to the multi-environment test and comparative analysis with foreign mainstream equipment, it shows that its technical indicators meet the requirements of ICAO Annex10. In this study, the full-link localization scheme of ILS signal transmitting unit from FPGA to RF front-end is initially realized, which provides practical reference for the autonomous and controllable technical path of aviation navigation equipment.

Abstract:The operational safety and health of key components of mechanical equipment are closely related to the temperature status. Aiming at the limitations of traditional single-point temperature measurement schemes that are difficult to capture the dynamic heat flow field, this paper proposes a multi-channel acquisition system based on thin-film resistance temperature detectors （RTD） sensors. The system aims to overcome the inherent defects of the on-resistance of the analog switch and the wire impedance on the accuracy constraints in the existing multi-channel acquisition. Innovatively, a three-wire system combined with a dual analog switch differential compensation architecture is complemented by a dual constant current source structure to effectively eliminate the effects of wire impedance and analog switch on-resistance. In addition, the polynomial fitting algorithm is utilized for real-time software compensation to further eliminate the resistance residual deviation. The experimental validation shows that the system can test the thin-film RTD sensors within the temperature zone of 20℃~300℃ with a resolution of 0.07℃ and an accuracy of better than 0.4℃, which is capable of providing highly reliable temperature field data for the structural health monitoring system of the key components, and improving the accuracy of the structural condition diagnosis of the mechanical components.

Abstract:In view of the lack of annotation standards in the current drainage pipe defect dataset, taking the pipe deformation defect as the research object, the YOLOv8 detection model was used to study the influence of different annotation methods on the pipe defect recognition results. Firstly, the defect images were classified into two major categories based on the characteristics of the deformation defects: Convex block deformation and overall deformation. For the convex block deformation, two labeling methods, namely convex block full marking and convex block half marking, were adopted. For the overall deformation, both overall full marking and overall half marking were used for labeling. Five independent tests were conducted for each labeling method. The results show that for the convex block deformation using the full labeling method of convex blocks, the values of the four evaluation indicators P, R, F1 and mAP50 are 76.56%, 74.36%, 75.43% and 72.12% respectively. The four values of the semi-labeled convex block annotation method are 88.51%, 76.77%, 82.17% and 83.14%, respectively. Therefore, the semi-labeled convex block annotation method is more suitable for convex block deformation. For the overall deformation, when the overall full labeling method was adopted, the values of the four evaluation indicators P, R, F1 and mAP50 were 83.85%,84.92%,84.34% and 84.46%, respectively. However, when the overall semi-labeling method was used for three training sessions, it was found that none of the four evaluation indicators could tend to stabilize. It indicates that the overall full-label annotation method is more suitable for overall deformation.

Abstract:Person re-identification, a core pillar of intelligent surveillance and smart city development, often exhibits significant accuracy degradation in real-world scenarios due to occlusion. Existing convolutional neural network-based person re-identification methods are constrained by local receptive fields, making long-range dependency capture across occluded regions difficult. Transformer-based person re-identification methods, despite global modeling capabilities, suffer from insufficient local-global feature fusion, leading to poor robustness in severely occluded scenarios. An end-to-end occluded person re-identification method based on dynamic feature enhancement and hierarchical gated fusion is proposed to tackle these problems. It employs a dynamic feature enhancement module to optimize local details and noise resistance of mid-level features, and a hierarchical multi-scale gated fusion module to mitigate semantic dilution in high-level features, constructing an end-to-end feature processing pipeline of "mid-level enhancement-high-level purification". The proposed method is compared with existing methods on Occluded-Duke, Occluded-ReID, Market1501 and MSMT17 datasets. Experimental results show Rank-1 accuracies of 74.8%, 88.8%, 96.7% and 90.9%, with mAP accuracies of 67.0%, 86.3%, 93.8% and 77.6%, respectively, validating its effectiveness and superiority in occluded scenarios.

Abstract:Aiming at the difficulty of occlusion and blur detection caused by complex detection scene in the process of wire harness terminal crimping, a wire harness terminal defect detection model ACS-YOLO based on improved YOLOv11 is proposed. The C2PSA_EFA module is designed in this model. Combining C2PSA with edge-enhanced feature attention, sobel operator is used to extract and fuse edge information to enhance the irregular defect capture ability of harness terminals. The attention scale sequence fusion module ASF-YOLO is introduced to improve the Neck part, and the multi-scale feature fusion mechanism is introduced to improve the detection ability of the model to the wire harness terminal defect features. The SlideLoss classification loss function is introduced to adjust the loss weight according to the difficulty of the sample, and the detection ability of the model to difficult samples is improved. The experimental results show that the accuracy, recall rate and mAP50 of the ACS-YOLO model are increased by 6.1%, 1.0% and 3.1% respectively compared with the original model, which can be effectively applied to the harness terminal defect detection task.

Abstract:YOLOv8 algorithm has been widely used in various target detection due to its high inference speed and excellent detection performance. However, when faced with small target detection and complex background interference in ship images, it faces many problems, such as missed detection, false detection and soon. In this paper, a RAMW-YOLOv8 target detection algorithm based on multi-scale features is proposed. By adding a convolution operation and a residual structure into the C2f module, and introducing the coordinate attention mechanism, a C3Res_CA module is constructed to enhance the ability of fine feature extracting and background noise suppression in complex background images; by integrating the multi-scale feature into the SPPF module, and introducing the average pooling layer and adaptive pooling operator to construct the SPPF_AuxPool module, which enhances the ability to mine different feature types. To solve the problem that small targets are dense and easy to interfere in ship images, a small target detection layer MicroDetect is added to enhance the detection ability for small size targets through multi-scale feature fusion and refined feature extraction strategy; to reduce the impact of low-quality samples on the accuracy of the algorithm, the WIoU loss function is introduced to increase the convergence speed, robustness and stability in complex scenarios. Experiments on the public dataset HRSID show that the RAMW-YOLOv8 algorithm improves the accuracy, recall and average precision of two different indexes by 1%, 3.6%, 3.1% and 3.2%, respectively, compared with the original algorithm, and its detection effect is obviously better than other classical algorithms.

Abstract:To address the challenges faced by wall-climbing robots during bridge inspections—namely low accuracy in identifying densely clustered, visually complex concrete defects across varying scales, leading to false positives and false negatives—we propose a lightweight CDM-YOLO algorithm based on YOLOv12n. First, to tackle the difficulty in recognizing multi-scale defects, we introduce a multi-scale feature fusion network into the backbone. This enhances the backbone′s ability to extract diverse and fine-grained features, enabling the model to adapt to defects of varying scales. Second, to address the confusion between dissimilar defects, a dynamic tanh mechanism is employed in the neck to enhance feature focus, clearly distinguishing different defects and reducing false positives and negatives. Finally, for densely clustered defects, the CARAFE algorithm is applied in the neck to strengthen deep semantic information flow, optimizing the model′s ability to identify dense defects. These methods improve detection accuracy while maintaining real-time performance and lightweight characteristics. Compared to YOLOv12n, CDM-YOLO achieves a 2.43% improvement in mAP (IoU=0.5) and a 3.25% increase in recall. This demonstrates its superior handling of multi-scale and dense defects with lower false positive and false negative rates, making it suitable for wall-climbing robots and field equipment with limited computational resources.

Abstract:Wireless powered mobile edge computing networks provide a promising solution for supplying both computational capability and stable energy to Internet of Things applications. However, compared with cloud environments, task arrivals and wireless channel conditions at the network edge exhibit stronger temporal dynamics, making fixed deployment schemes insufficient for adapting to workload variations. To address this issue, this paper considers a multi-charger Wireless powered mobile edge computing network and formulates an optimization problem that maximizes the computation completion rate of wireless devices by jointly optimizing wireless charger online deployment decisions, task offloading ratios, and resource allocation. To solve this mixed-integer non-convex problem, the original problem is decomposed into a wireless charger deployment decision subproblem and a resource allocation subproblem. Then, a Gated Transformer-based joint optimization algorithm is proposed to effectively handle long-term network dynamics and high-dimensional action spaces. In this work, wireless charger deployment is modeled as an online discrete decision over a predefined set of candidate locations, where the deployment scheme is updated frame by frame to adapt to network dynamics. Simulation results show that, compared with baseline algorithms, the proposed method improves the average computation completion rate by about 48% under normal workloads and by up to 60% in high-load computationintensive scenarios, while maintaining good stability and convergence.