Abstract:In order to solve the problems that it is difficult for service robots to accurately understand pedestrian intentions and unreasonable obstacle avoidance path selection in dynamic pedestrian environment, a pedestrian openness comfort model was proposed. Firstly, by extending the traditional two-dimensional symmetric Gaussian function to an asymmetric Gaussian function, the dynamic comfort space of pedestrians can be modeled more accurately. Secondly, combined with the pedestrian′s head posture and pedestrian′s openness characteristics, the robot′s ability to understand the pedestrian′s movement intention and social interaction relationship is enhanced, so as to improve the friendliness and rationality of navigation. Finally, through the comparison and verification of simulation and experiments in the real environment, the service robot using the pedestrian openness comfort model is more optimized in path selection, and can actively avoid the interactive space of pedestrian groups, which not only reduces the possibility of conflict with pedestrians, but also enhances the smoothness and naturalness of navigation, and shortens the movement time by 1.15 and 2.58 s respectively in the simulation environment of different scenes. In the real environment of different scenes, the exercise time was shortened by 1.14, 2.30 and 0.12 s, respectively. Experimental results show that the model can effectively adapt the robot to complex pedestrian dynamic scenes, improve the efficiency of obstacle avoidance, and significantly improve the social friendliness and navigation quality of the robot in the human-machine integration scene.

Abstract:In order to solve the problems of power and frequency fluctuations and harmonic content in the output voltage occurring in the grid-connected PV under the traditional virtual synchronous generator control, a VSG rotational inertia adaptive control method and a modulation scheme with stochastic excitation are introduced in the grid-connected. A voltage control loop with virtual impedance is introduced in the VSG control and combined with a current control loop based on a quasi-proportional resonant controller to construct a VSG control strategy for grid-connected inverters of PV power systems. With this strategy, the THD of the three-phase voltages A, B and C decreased by 15.17%, 15.37% and 13.10%, respectively, and the active power overshoot decreased by 7.42% in simulation results, and the THD of the three-phase voltages A, B and C decreased by 1.92%, 4.61% and 2.44%, respectively, in experimental results, and the frequency was stabilized at 50.07 Hz. The simulation and experimental results demonstrated that the proposed method can effectively suppress the power and frequency oscillations and reduce the THD of output voltage, which verifies the feasibility of the proposed method.

Abstract:In the logistics robot transportation process, path planning is the core link, facing challenges such as insufficiently smooth paths and low algorithm search efficiency. The A* algorithm, as a widely used global path planning method, has problems such as ineffective path smoothing when applied to logistics robots. To this end, the traditional A* algorithm has been improved by dynamically weighting the heuristic function and using the Floyd algorithm to remove redundant points in the path, while introducing a safe distance mechanism to prevent collisions. In addition, the path has been smoothed and optimized to better adapt to the actual movement needs of logistics robots. The MATLAB simulation results show that the improved A* algorithm reduces the average number of turning points by 58.5%, shortens the path length by 3.19%, and reduces the number of traversal points by 59.9% compared to traditional algorithms. Further combining with DWA algorithm for local path planning, obstacle avoidance function has been achieved. The effectiveness of the fusion algorithm has been verified through simulation and real vehicle experiments.

Abstract:In order to address the issues of missed detections, false positives, and low accuracy in small traffic sign detection, this paper proposes a detection model for small traffic signs, named YOLOv8-Faster-Ghost-GAM. The algorithm introduces a global attention mechanism (GAM) into the last C2f module of the backbone network, enhancing key features and suppressing irrelevant information to significantly improve the detection of small targets and the recognition capability in complex scenarios. Additionally, each C2f module in the backbone network is replaced with FasterNet to reduce the number of model parameters, and standard convolutions are replaced with Ghost convolutions, which use inexpensive linear transformations to reduce computational effort. Finally, the WiOU loss function is employed to effectively improve the recognition of low-quality samples, resulting in a 1.6% increase in precision and a 3.2% increase in recall, thereby demonstrating the effectiveness of the proposed improvements.

Abstract:Most existing dynamic simultaneous localization and mapping (SLAM) algorithms simply remove dynamic objects, resulting in the loss of dynamic object motion information that aids in the system′s own localization and navigation, and have limitations for complex and ever-changing industrial environments. In this paper, we propose an improved visual SLAM algorithm for target tracking that performs localization while obtaining a more accurate estimate of the object′s pose. The algorithm uses background points for its own localization, uses refined optical flow information to reduce the effect of noise for accurate localization, and then combines the scene flow information with polynomial residuals to obtain accurate dynamic object sensing results and to reduce the algorithm′s error in estimating the object′s pose. Finally, the proposed algorithm is evaluated on the publicly available KITTI Tracking dataset and real scenes. The experimental results show that on the public dataset, the proposed algorithm has an average rotation error (RPER) of 0.027° and an average displacement error (RPET) of 0.069 m. The average rotation error of object pose estimation is 0.686 97°, and the average displacement error is 0.103 50 m. The proposed algorithm is able to have a better performance of self-localization and dynamic object tracking. The proposed algorithm also shows excellent localization and tracking performance in real scenarios.

Abstract:In order to solve the large pose estimation error of radiant field Visual SLAM algorithm and poor robustness in the process of fusion with inertial measurement unit, this paper proposes a radiance field visual inertial SLAM algorithm based on tightly coupled IMU. The algorithm uses an improved pre-integration module to implement a tightly coupled framework, the improved initialization strategy to deal with the robustness problem, combined with radiation field loss to optimize pose and bias. The proposed algorithm is applied to the positioning modules of NICE-SLAM and MonoGS, and is experimentally tested on the IMU-RGBD dataset OpenLORIS, and the tight-coupled module can improve the positioning accuracy by 34.3% and 14.8% respectively. Compared with MM3DGS, the proposed algorithm has higher robustness, which can effectively improve the positioning accuracy and has a good generalization ability to improve the SLAM performance of the radiance field.

Abstract:A lightweight and efficient bearing defect detection algorithm DWA-YOLO is proposed to address the challenges of large scale variation, similar texture, and dense distribution of defects in the surface defect detection of bearing outer rings, as well as the complexity of existing detection model structures, poor computational complexity, and detection accuracy. Firstly, a plug and play lightweight dual bottleneck structure module DBM was designed to effectively reduce model complexity and enhance the model′s ability to extract features at different scales. Secondly, the wavelet convolution WTConv with multi-scale characteristics is introduced as a downsampling operator in the network backbone. By expanding the receptive field of the model and utilizing the multi-scale analysis characteristics to capture the details and texture information of the image, the model′s anti-interference ability against texture and noise and its ability to understand contextual information are enhanced, thereby improving the overall detection accuracy. In addition, this article designs a joint loss function Alpha-MPDIOU, which utilizes power transformation mechanism to improve the localization accuracy of bounding boxes and solve the problem of detecting multiple boxes. Finally, the use of auxiliary detection head training strategy accelerates the convergence speed of the model and enhances its detection capability. The experimental results show that DWA-YOLO improves mAP accuracy by 3.5% compared to the baseline model, with a model parameter size of 2.6 M and a computational complexity of 7.4 GFLOPs. The improved model not only enhances the ability to identify bearing defects, but also reduces network complexity, making it more suitable for the detection needs of bearing outer ring surface defects in industrial sites.

Abstract:In order to solve the problems of the traditional threevector model predictive control strategy, such as large calculation amount of vector selection, complicated calculation of operation time of each vector and large common mode voltage, a multi-vector model predictive current control strategy is proposed. Firstly, to solve the problem of large common-mode voltage, it is proposed to replace traditional zero vector with effective voltage vector synthesis, and use voltage vector selection table and voltage vector position angle to quickly select vectors and reduce the calculation amount of vector selection. Secondly, voltage error duty cycle is adopted to simplify the calculation of the operation time of each vector. Finally, its effectiveness is verified by simulation and physical platform. It is proved that the control algorithm can improve the steady-state performance of the system and restrain the influence of large common mode voltage on the motor.

Abstract:An improved Informed-RRT* algorithm is introduced to tackle issues related to high randomness, a large number of infeasible nodes, and low convergence efficiency in path planning. This algorithm optimizes node usage through global sampling and an adaptive step size. The initial path is generated using a biased bidirectional search and a parent node reselection technique, which offers a more effective starting point for further iterative optimization. During the elliptic iteration, a greedy approach is applied to eliminate unnecessary nodes. Additionally, path backtracking is refined to decrease redundant nodes and improve trajectory smoothness. This study presents two factors: obstacle complexity and map size, to assess the performance of the enhanced algorithm against the original Informed-RRT* algorithm in four different scenarios. Results from 20 experiments show that the improved algorithm decreases the number of trajectory waypoints by 28.6% to 64.3% and reduces trajectory length by 0.3% to 2.7%. These results suggest that our enhanced method enhances node utilization, produces shorter trajectories, and significantly cuts down on computational iterations compared to the Informed-RRT* algorithm.

Abstract:Sonar detection technology has been widely used in underwater structure detection. Affected by the complex underwater environment, sonar images usually have substantial problems such as low resolution, serious noise interference, fuzzy edge details, and poor texture information. In order to solve these problems, this paper proposes a fusion denoising algorithm based on improved anisotropic guided filtering and Wiener filtering. Firstly, the local structural similarity index was introduced into the traditional AnisGF as a weighting factor to achieve denoising while retaining more edge structure information. Secondly, the Bayesian optimization method was used to determine the SSIM weight of Wiener filtering. Finally, AnisGF and Wiener filtering were combined for joint denoising of sonar images. The experimental results show that the proposed algorithm has 9.5%, 4% and 10% improvements in mean square error, peak signal-to-noise ratio and structural similarity index compared with the traditional algorithm.

Abstract:Spectrum sensing is one of the key technologies to alleviate spectrum resource shortages, and intelligent spectrum sensing has become a hot research direction. To address the issues of insufficient feature extraction in existing spectrum sensing methods and poor sensing performance under low signal-to-noise (SNR) ratio conditions, a hybrid spectrum sensing model is proposed. The model consists of an Inception module, bidirectional gated recurrent unit, temporal attention mechanism, and fully connected layer network. Firstly, the Inception module extracts multi-scale spatial features from the received I/Q signals. Then, the bidirectional gated recurrent unit is used to capture the temporal sequence features of the signals, while the temporal attention mechanism enhances important temporal features. Finally, the fully connected layer network maps the extracted features to the classification space of spectrum states to complete classification and recognition. The experimental results show that the proposed method significantly improves perception performance compared to several existing spectrum sensing methods. The overall detection accuracy of the model reaches 84.55%, and when the SNR is -20 dB, the perception error of the method is 24%. The proposed method also demonstrates good adaptability to various modulation types of radio signals. It does not rely on any prior information and exhibits strong robustness in low SNR and complex radio environments. This approach achieves an effective balance between perception performance and model complexity, providing a new solution for intelligent spectrum sensing.

Abstract:Intelligent reflecting surface (IRS) is one of the key technologies in the sex generation(6G). However, for multi-user systems, the computational complexity of the system increases greatly with the increase of the number of reflective units and the number of users, and the optimal design of the system faces great challenges. In this paper, we propose a low computational complex transmission rate maximization algorithm based on multi-user reflection unit selection. According to the user′s rate requirements and channel conditions, the algorithm selects the matching reflection unit, considers the phase shift setting and the base station beamforming, and carries out joint optimization to establish a user rate maximization problem. There is a high degree of coupling between the variables in this optimization problem. Therefore, the original problem is divided into two subproblems for solving, and the approximate solution is obtained by using semidefinite relaxation. The simulation results show that the algorithm proposed in this paper can significantly reduce the computational complexity of the system while improving the downlink transmission rate. Compared to a system without IRS assistance, the transmission rate increases by about 50%; compared to a random phase IRS, the transmission rate increases by about 30%.

Abstract:To investigate the effect of different guidance methods on the cortical activation during fine motor imagery, a novel fine motor imagery method combining visual and auditory guidance is proposed. The goal is to explore the enhancement of cortical activation during fine motor imagery using different guidance approaches and to uncover the underlying patterns of cortical activity. An experimental paradigm was designed for fine motor imagery of the wrist, elbow, and shoulder joints, incorporating four guidance methods: simple visual guidance, auditory guidance, dynamic visual guidance, and dynamic visual combined with auditory guidance. ERD and ERS in the time and frequency domains were used as metrics to assess cortical activation. Energy distribution and brain network functional connectivity were utilized to observe the spatial distribution of cortical activity and analyze the degree of cortical activation under different guidance methods. The experimental results indicate that the dynamic visual combined with auditory guidance led to significantly higher ERD and ERS amplitudes compared to the other guidance methods. Additionally, under the visual and auditory combined guidance, the activated cortical regions were more extensive, and stronger synchronization and desynchronization were observed in multiple brain areas. Compared to simple visual guidance, auditory guidance, and single dynamic visual guidance, the dynamic visual combined with auditory guidance significantly enhanced cortical activation during fine motor imagery. This method provides a new guidance strategy for fine motor imagery training, contributing to improved training effectiveness and rehabilitation efficiency, with potential practical applications.

Abstract:In the BeiDou satellite synchronization system, FPGA-based solutions are typically used. However, using an ARM single-core system during scheduling can lead to resource contention and real-time response deviations. While ARM processors are superior to FPGA in handling business logic, floating-point calculations, and similar tasks, this paper proposes a solution for BeiDou 1PPS synchronization and timing based on ARM processors. The synchronization calculation is implemented using the least squares method combined with a sliding window, while the timing calculation is achieved through a phased growth mechanism. Additionally, a delay correction algorithm is introduced to address cycle boundary acquisition deviations caused by interrupt conflicts during signal processing. When the system detects that the data is about to overflow, the algorithm delays recording the rising edge signal′s cycle value and applies corrections. Experimental results show that this algorithm can achieve synchronization accuracy at the level of 10-8 s, proving its effectiveness in high-precision time synchronization applications.

Abstract:Addressing the challenges in the domain of image inpainting, such as the high computational complexity, loss of information during feature extraction, and the blurring of textures in the inpainting images, this study proposed a image inpainting model that integrates multiscale hierarchical feature fusion with synergetic global-local Transformer. Initially, the multi-scale hierarchical feature fusion block was proposed as a means of effectively fusing deep and shallow features in detail, thereby reducing the loss of key information while expanding the sensory field. Subsequently, synergetic global-local Transformer blocks for global reasoning was proposed, featuring an integrated rectangle-window self-attention mechanism and local feed-forward neural networks. This design reduced computational complexity while enhancing the model′s macroscopic understanding of global context and microscopic grasp of local detail characteristics.The proposed method was validated on the CelebA-HQ and Places2 datasets, and the results demonstrated that it yielded improvements in PSNR by an average of 0.26~6.25 dB, SSIM by an average of 1.4%~19%, and L1 decreased by an average of 0.2%~5.66% compared to commonly used inpainting methods when dealing with 40%~50% masks. The experiments show that the inpainted images resulting from the proposed method exhibit a more realistic and natural visual effect, thereby providing further validation of the method′s effectiveness.

Abstract:Aiming at the problems of high detection false alarm rate, low detection accuracy, leakage and false detection caused by dense target arrangement, large scale difference and complex background of remote sensing images, a remote sensing image detection algorithm YOLOv8-EP based on YOLOv8n is proposed. Firstly, a feature focus diffusion pyramid network (FFDPN) is constructed to capture multi-scale information through parallel deep convolution, while adding a diffusion mechanism to diffuse the feature information to each detection scale to enhance feature interaction. A lightweight task align dynamic detection head (TADD) is designed to improve the localisation and classification performance of detection through feature sharing and parallel task processing. Then, the SimAM attention mechanism is introduced to capture key information in the image and increase the model sensory field. Finally, the Inner-CIoU loss function is introduced to improve the detrimental effect of low-quality images on the network gradient and accelerate the model convergence. Experimental results on the NWPU VHR-10 dataset and RSOD dataset show that YOLOv8-EP achieves a mAP of 97.6% and 97.9%, respectively, with a 13% decrease in the number of parameters, and improves by 2.2% and 1.5% compared to the YOLOv8n baseline network, which can meet the requirements of industrial deployment and achieve good detection performance overall.

Abstract:Segmentation of coronary arteries is crucial for the rapid diagnosis of cardiovascular diseases. Given the challenges posed by the complex structure of coronary arteries and the interference from other vascular tissues, which often result in fragmented segmentation, ensuring the model′s ability to adapt to segmenting different morphological structures of the coronary artery, a novel 3D coronary artery segmentation network (CA-SegNet) is proposed. This model incorporates a combination of CNN and Transformer as the encoder and decoder, leveraging their advantages and complementarity to fully extract both global and local features of coronary arteries. By proposing a multi-scale feature interaction module, the model simultaneously extracts multi-scale features of coronary arteries while facilitating feature channel interaction. In the decoding stage, an attention weighted feature fusion module is proposed to weight and fuse features from both spatial and channel perspectives, enabling the model to focus more on the coronary artery regions. Experimental results demonstrate that the proposed model achieves DSC, Recall, Precision, and HD95 values of 81.96%, 84.24%, 80.11% and 14.94 respectively, surpassing current popular segmentation models and validating the effectiveness of CA-SegNet.

Abstract:In industrial settings, the acquisition and annotation of defective workpieces pose significant challenges, severely hindering defect detection efforts. While generating a large number of defective samples from limited real-world samples effectively mitigates the issue of sample scarcity, existing defect generation methods are often constrained by suboptimal visual authenticity and poor alignment with defect masks. To address these limitations, this study introduces AnomalyAlign, a novel controllable diffusion model designed to synthesize highly realistic industrial defect images with precise mask alignment. Leveraging the foundational knowledge of the text-to-image model Stable Diffusion, AnomalyAlign incorporates a semantic-aligned text prompt generator to produce text prompts that achieve closer semantic alignment with real images, thereby accelerating model convergence. Furthermore, the model integrates a defect alignment loss function, which enhances the spatial consistency between generated defect images and their corresponding masks. Extensive experimental validation on the MVTec-AD dataset demonstrates that AnomalyAlign generates defect images with superior realism and diversity, while significantly improving the performance of downstream defect detection tasks.

Abstract:In the tree barrier clearance project for protecting the safety of the electrical distribution network, the manual calculation of the felling quantity faces problems such as strong subjectivity of the calculation results and management difficulties. The existing algorithms have low accuracy, with many false positives and false negatives, and poor robustness. Therefore, a tree stump detection algorithm for calculating the felling workload in the transmission corridor tree barrier clearance is proposed. In response to the problem of inaccurate felling quantity calculation due to the complexity of the distribution network clearance scene and the difficulty in distinguishing between tree trunks and tree stumps, a feature extraction module based on Context Guide Block is designed. RepGFPN and Dysample structures are introduced to optimize the neck network, effectively integrating environmental context semantic information with local details of tree stumps. Subsequently, the algorithm designs a tree stump detection head based on LW-SEAM, optimizing the detection effect under occlusion. The model′s P, R, and mAP50 indicators on the test set have been improved to 85.5%, 76.4%, and 80.4% respectively, showing good detection performance for tree stump detection in complex backgrounds and occlusion scenarios, and providing technical reference for achieving intelligent engineering calculation.

Abstract:The installation of photovoltaic generation (PVG) systems in existing office buildings (EOB) is one of the environmental green energy measures. However, the fluctuation of PVG negatively impacts the stable electricity use of EOB, making EOB-PVG power prediction crucial. This paper proposes the EOB-PVG power prediction method using sparrow search algorithm-long short-term memory (SSA-LSTM). The method preprocesses the collected environmental and generation data using multiple imputation + principal component analysis (MI+PCA) and splits the dataset. The LSTM neural network prediction model is designed, and SSA is used to automatically optimize the neural network's hyperparameters to achieve accurate prediction. The experiment selects real environmental and generation data from the EOB, and after preprocessing, the cumulative contribution rate of the principal components of the dataset exceeds 95%. Three evaluation metrics are designed to assess prediction performance. Comparison results show that SSALSTM outperforms LSTM and SSA-TCN in prediction accuracy and fitting ability, providing good accuracy in predicting EOB-PVG power and contributing to the subsequent realization of intelligent energy management tasks for EOB.

Abstract:The tension clamp plays the role of connecting wires and carrying current in the transmission line, and its crimping quality is directly related to the safe and effective operation of the power grid. In order to solve the problems of complex operation and high personnel requirements in the DR defect detection of tension clamp crimping, a DR image defect evaluation method using VA-UNet segmentation technology was proposed. Firstly, the semantic segmentation model VA-UNet for DR image defects in tension clamps is studied, VGG16 with significant image feature extraction and analysis ability is selected as the backbone network, multi-scale feature fusion is enhanced by integrating spatial pyramid pooling structure ASPP, and mixed loss function is introduced to accelerate the model convergence and improve the segmentation accuracy. Then, a grading method combining the model prediction segmentation results and related quantitative analysis was studied to realize the hazard severity assessment of DR defects in tension clamp crimping, which provided a reference for the subsequent wire clamp treatment. Based on the data set preparation and the analysis of test evaluation indicators, the relevant ablation experiments showed that the mIoU and mPA of VA-UNet reached 84.14% and 91.58%, respectively, which were significantly higher than those of the original model. The experiment of assessing the severity of DR defects in tension clamp crimping shows that the method is scientific and practical.

Abstract:In semantic communication, image semantic information processing heavily relies on computationally intensive convolutional neural networks, which require higher computational performance, especially when handling high-resolution images. This presents a significant challenge for the application of semantic communication in edge scenarios. To address this, this paper proposes an FPGA-based semantic information processing accelerator, which innovatively integrates the convolutional neural network encoder and rANS encoding in the same hardware accelerator. Specifically, the accelerator adopts a systolic array architecture combined with multiplyaccumulate units, loop tiling strategy, and a dual-buffer structure to fully leverage the parallel computing capabilities and on-chip storage resources of the FPGA, improving data transmission efficiency and computational performance. Each processing unit integrates multiple multiply-accumulate units, capable of performing two INT8 multiplications and local accumulation in each clock cycle. Finally, rANS is used for 8-way parallel encoding of the output features, further compressing the feature data. Experimental results show that, on the ZCU104 platform, the design achieves a throughput of 300.5 GOPS with a power efficiency of 66.77 GOPS/W when processing 1080P images, providing a processing speed approximately 6 times faster than Intel CPUs and 58 times faster than ARM CPUs. Compared with other FPGA accelerators, the BRAM efficiency improves by approximately 730%, 40%, and 63%, the energy efficiency by approximately 802%, 60% and 3%, and the DSP efficiency by approximately 476%, 70% and 133%. The proposed accelerator demonstrates significant performance advantages and can efficiently process image semantic information, offering broad practical application potential.