Abstract:In view of the weaknesses of poor computing and storage capabilities of edge devices, lightweight processing was carried out on the backbone network CSPDarkNet53 for feature extraction in the traditional YOLOv5 model, and a lightweight gesture recognition algorithm MPE-YOLOv5 was proposed to realize the deployment of the model in low-power edge devices. Considering the problem that it is difficult to identify large-scale transformation targets and tiny targets due to less feature extraction in lightweight model, efficient channel attention (ECA) mechanism was added to alleviate the loss of information after high-level feature mapping due to the reduction of feature channel. A detection layer for tiny targets was added to improve the sensitivity to tiny target gestures. EIoU was selected as the loss function of the detection frame to improve the positioning accuracy. The effectiveness of the MPE-YOLOv5 algorithm was verified on the self-made dataset and NUS-Ⅱ public dataset, and the MPE-YOLOv5 algorithm was compared with lightweight M-YOLOv5 algorithm and original YOLOv5 algorithm on the self-made dataset. Experimental results show that the model parameters, model size, and computational complexity of the improved algorithm were 21.16%, 25.33%, and 27.33% of the original algorithm, and the average accuracy was 97.2%. Compared with the lightweight model M-YOLOv5, MPE-YOLOv5 improved the average accuracy by 8.72% while maintaining the original efficiency. The proposed MPE-YOLOv5 algorithm can better balance between the detection accuracy and real-time reasoning speed of the model, and can be deployed on edge terminals with limited hardware.

Abstract:Due to the complex background of ship targets and much irrelevant interference in visual images, it is difficult to conduct ship detection. In addition, there are few datasets for multi-category ship detection and the samples are often unbalanced, which makes the ship target detection performance degraded. Considering the ship detection background interference, an improved YOLOv3 model was proposed by introducing SimAM attention mechanism, which was used to enhance the weight of the ship target in the extracted features and suppress the weight of background interference, thus improving the model detection performance. Meanwhile, strong real-time data augmentation was applied to improve the unbalanced distribution of sample scales, and transfer learning was combined to improve the ship detection accuracy in the condition of a restricted number of samples. The visualization results of extracted features show that the improved model could suppress irrelevant background features, and the abilities of feature extraction and target localization were enhanced. Without introducing additional learnable parameters, the proposed model achieved 96.93% and 71.49% for mAP.5 and mAP.75 on the SeaShips dataset, and detection speed reached 66 frames per second, indicating a good balance between detection accuracy and efficiency. The improved model optimized the target features more effectively compared with the Saliency-aware CNN and eYOLOv3 models, resulting in an improvement of mAP.5 by 9.53% and 9.19%. The mAP.5 for ship type target detection on Singapore Maritime Dataset reached 81.81%, indicating that the proposed model has good generalization performance.

Abstract:In view of the problem of high complexity of the shaped-offset quadrature phase-shift keying (SOQPSK) symbol timing algorithm under high-precision conditions, an optimized maximum likelihood (ML)-based SOQPSK joint symbol timing and phase recovery algorithm was proposed. The SOQPSK-MIL signal model and the phase changing regulation of the signal were presented. The principle and estimation results of the ML-based algorithm for SOQPSK signal were given. The energy difference between different h functions in the algorithm was analyzed in detail, and the performance of the algorithm under different simplification degrees was discussed. The effectiveness of the algorithm when it only used Z+₁ and Z+_－1 parameters for estimation was analyzed. Considering the problem of the large amount of calculation during the convolution process in the implementation of the algorithm, an optimized processing method of h function was proposed. On the premise of ensuring that the energy of the h function remained unchanged, an optimization method was proposed, which quantizes the h function to make it a simple three-valued function with only three amplitudes of 0, A, and -A, thereby effectively reducing the calculation amount of the convolution process and reducing the computational complexity of the convolution part from O(N²L₀) to O(NL₀). Simulation results show that the algorithm had the characteristics of high estimation accuracy and low computational complexity. The implementation complexity of the algorithm was greatly reduced, and the previous high estimation accuracy was maintained. The minimum mean square errors of the timing error and phase error estimation results of the improved algorithm were almost the same as those before the improvement.

Abstract:With the increasing scale of network, the accurate and real-time prediction of network flow is essential for traffic scheduling and routing design. However, due to the nonlinearity and uncertainty of network flow data, some traditional methods fail to achieve good prediction accuracy. Considering the complex spatialtemporal features of network flow, a novel network flow prediction method based on spatialtemporal features fusion (ST-Fusion) was proposed, combined with encoderdecoder architecture. First, the encoder was designed with two parallel feature channels: temporal and spatial channels. The temporal features were extracted by integrating gated recurrent unit (GRU) and self-attention mechanism, and the graph convolutional network (GCN) was used to extract the spatial features. Then, the temporal and spatial features extracted by the encoder were fused by using the bilateral gated mechanism. Finally, the fused features were input into the GRU-based decoder to generate prediction results. Experiments were conducted on three public datasets (GEANT, ABILENE, and CERNET) using evaluation metrics including MAE, RMSE, ACCURACY, and VAR. Experimental results showed that the ST-Fusion method achieved better performance in network flow prediction.

Abstract:The wear of cutter is an important factor affecting the efficiency of shield tunneling, which is also a basis for determining the time and frequency of cutter replacement. As it is difficult to evaluate the overall wear state of the cutters in the process of shield tunneling, three wear degrees (light, moderate, and severe) were proposed based on the relationship between the wear amount of each cutter and the limited wear at the cutters change site. The theoretical relationship between three main tunneling parameters (thrust, torque, and tunneling speed) and the cutting force component of a single cutter was derived, and a method for recognizing the overall wear state of cutters was proposed by using the wavelet packet algorithm to decompose the tunneling parameter signals. In this method, the eigenvectors composed by the standard deviation of the wavelet packet coefficient of decomposed signal nodes were used as the wear recognition index. Sensitivity analysis was performed to find out the most sensitive node eigenvector of the cutter wear. Then the functional relationship between the wear state and the wear recognition index was determined by fitting. The analysis of the section from Dayun station to Baohe station of Shenzhen Metro Line 14 showed that the method could accurately recognize the overall wear state of the shield cutters. Among the three tunneling parameters, the recognition accuracy was the highest when using the tunneling speed signal, followed by the thrust signal, and the torque signal was the lowest. The proposed method is easy to use and cost-effective, since it only needs to analyze the automatically collected tunneling signals without installing any sensors, which provides reference for cutter replacement.

Abstract:In view of the problem that general pre-trained models are not suitable for named entity recognition tasks in the medical domain, a neural network architecture that integrates knowledge graph in the medical domain was proposed. The elastic position and masking matrix were used to avoid semantic confusion and semantic interference in self-attention calculation of pre-trained model. The idea of multi-task learning in fine-tuning was adopted, and the optimization algorithm of recall learning was employed for pre-trained model to balance between general semantic expression and learning of the target task. Finally, a more efficient vector representation was obtained and label prediction was conducted. Experimental results showed that the proposed architecture achieved better results than the mainstream pre-trained models in the medical domain, and had relatively good results in the general domain. The architecture avoided retraining pre-trained models in particular domain and additional coding structures, which greatly reduced computational cost and model size. In addition, according to the ablation experiments, the medical domain was more dependent on the knowledge graph than the general domain, indicating the effectiveness of integrating the knowledge graph method in the medical domain. Parameter analysis proved that the optimization algorithm which used recall learning could effectively control the update of model parameters, so that the model retained more general semantic information and obtained more semantic vector representation. Besides, the experimental analysis showed that the proposed method had better performance in the category with a small number of entities.

Abstract:Image segmentation is to divide the region with special meanings into several disjoint sub-regions according to certain rules, which is the key link between image processing and image analysis. The traditional watershed image segmentation method is widely used, which has the advantages of fast and simple. However, it is easily interfered by noise, and the segmentation results are prone to lose important edge information, resulting in over-segmentation. In view of the problem of the traditional watershed image segmentation method, an improved watershed image segmentation method based on adaptive structural elements was proposed. First, the adaptive structural elements with variable shapes were constructed by using local density, symmetry, and boundary features of adjacent pixels of image targets, so as to ensure a good consistency between the proposed structural elements and the shape of image targets. Then, the adaptive structural elements were used to obtain the morphological gradient of the image, which could improve the positioning accuracy of the target edge. The L₀ norm gradient minimization and morphological open-close hybrid reconstruction were used to modify the gradient image, so as to reduce the local invalid minimum points in the gradient image and suppress the occurrence of over-segmentation. Finally, watershed segmentation was performed on the modified gradient image to realize accurate segmentation of the target region of the image. Experimental results show that the method could effectively restrain over-segmentation of traditional watershed algorithm and improve the accuracy of the target edge positioning, with high precision of image segmentation.

Abstract:Considering that the interference suppression capability of the existing adaptive beamformer decreases when coupling the array element position error, amplitude and phase error, and signal arrival direction error, a robust adaptive beamforming algorithm based on eigenspace bases transition was proposed. First, the influence of the array element position error, amplitude and phase error, and signal arrival direction error of the steering vectors was modeled. Then, on the basis of the characteristics that the real signal subspace was the same as the space spanned by the steering vectors, the concept of subspace distance was introduced to quantify the similarity of two subspaces, and a multi-dimensional nonlinear optimization problem that minimized the spatial distance was constructed. Next, a hybrid optimization strategy was formed by combining the characteristics of the genetic algorithm and the quasi-Newton method, and after solving the optimization problem, a set of non-orthogonal bases of the signal subspace were obtained. Finally, the estimated signal subspace and the noise subspace were combined into an eigenspace. The accurate interference-plus-noise covariance matrix was extracted by the bases transition of the eigenspace, and the steering vector of the desired signal was corrected. Numerical simulation results show that the proposed hybrid optimization algorithm could significantly reduce the subspace distance as the number of iterations increased. When the number of iterations reached 100, the subspace distance reduced to less than 1. When the array element position error, amplitude and phase error, and signal arrival direction error occurred at the same time, and the input signal-to-noise ratio was 10 dB, the output signal-to-interference noise of the proposed algorithm was about 14 dB higher than that of the existing methods.

Abstract:Considering the failure mechanism and weaknesses of traditional fabricated shear wall structures under strong earthquakes, a new type of fabricated shear wall with functions of energy dissipation and shock absorption was proposed. On the basis of model test and numerical simulation, seismic performance tests were carried out on four specimens with scale ratio of 1∶1.54 and shear span ratio of 1.52. Further analysis was conducted to investigate the effects of bolt number, axial compression ratio, and reinforcement ratio of edge members on the seismic performance of the new fabricated shear wall, including failure modes, hysteretic performance, bearing capacity, displacement ductility, stiffness degradation, and energy dissipation capacity. Test results show that the four specimens experienced shear compression failure, which was the same as the cast-in-place shear wall with the same shear span ratio. However, the proposed shear wall had better hysteretic performance and energy dissipation capacity, and the energy dissipation capacity was higher than that of the cast-in-place shear wall at the failure point. When the number of bolts decreased, the hysteretic performance of the new fabricated shear wall decreased, the wall deformation increased, while the bearing capacity remained almost unchanged. When the axial compression ratio or reinforcement ratio of edge members decreased, the bearing capacity decreased, and the ultimate displacement increased. Finally, the finite element model of the specimens was established by ABAQUS program. Comparisons of numerical results and test results showed a good agreement, verifying the correctness of the model, which can be applied to the analysis of the new fabricated shear wall.

Abstract:In view of the problems of low brightness and obvious color distortion of the sky in restored images in most existing algorithms for image dehazing, a haze removal method for UAV aerial images based on atmospheric light value and graph estimation was proposed. First, the depth-of-field image was obtained according to the color attenuation prior theory, and the mean value of the region with the minimum deviation in the depth-of-field image was taken as the atmospheric light value. Then, a random walk clustering method was designed to estimate the atmospheric light map. The random walk algorithm was used to cluster the image into N sub-regions, and the mean value of the first 0.1% pixels of the sub-regions was taken as the regional atmospheric light value, which was then combined and refined by guided filtering to obtain the atmospheric light map. Next, the two atmospheric light estimators were fused into a new atmospheric light map with atmospheric light valuegraph estimation, which is a more accurate atmospheric light estimator. The transmittance was obtained by haze-lines prior method, and a dark compensation method was proposed to improve the transmission accuracy. Finally, according to the atmospheric scattering model, a clear restored image was obtained based on the fused atmospheric light map and optimized transmittance. Experimental results show that compared with other algorithms, the proposed algorithm improved the information entropy, mean gradient, blur coefficient, and contrast by 1.1%, 6.3%, 8.5%, and 6.4%, respectively, with better subjective visual effect and more abundant information.

Abstract:To update the structural finite element model through stochastic static displacement measurement data and maintain the computational efficiency, we proposed a stochastic model updating method based on homotopy meta-model and Bayesian sampling method. First, the objective function was constructed by using the static displacement of the structure, and the delayed rejection adaptive sampling algorithm was used to estimate the posterior probability density of the updated parameters. In the process of sampling, the homotopy meta-model was adopted instead of the finite element model to calculate the static displacement of the structure. Numerical examples and test results show that when updating the finite element model of variable cross-section beam, as opposed to the quadratic response surface model, by incorporating the homotopy meta-model into the static Bayesian model method, the posterior probability density of the updated parameters could reproduce the stochastic response of the structure more accurately, making the probability density function of the stochastic response of the updated structure more consistent with that of the measured results. Even when the coefficient of variation of the stochastic measurement error was large and the difference between the prior information and the real updated parameters was large, the proposed method could quickly obtain the posterior probability density of the updated parameters, so that the probability density function of the structural stochastic displacement response calculated by the updated parameters was consistent with that of the measured results. The homotopy meta-model combined with Bayesian sampling algorithm can update the stochastic model of the structure quickly and accurately within the probability framework.

Abstract:In the ultra-dense network environment, each access point is deployed in the hotspot area, which forms a complex heterogeneous network. Users need to choose the appropriate network to access, so as to achieve the best performance. Network selection problem is to choose the optimal network for the user, so that the user or network performance reaches the best. In order to solve the access selection problem of users in ultra-dense networks, we proposed an ultra-dense network access selection algorithm based on the improved deep Q network (DQN), considering network states, user preferences, and service types, and combining with load balancing strategies. First, by analyzing the influence of network attributes and user preferences on network selection, the appropriate network parameters were selected as the parameters of the access selection algorithm. Then, the problem of network access selection was modeled by Markov decision-making process, and the states, actions, and reward functions of the model were designed. Finally, the optimal network strategy was obtained by using DQN to solve the network selection model. In addition, the target function of traditional DQN was optimized to avoid overestimation of Q value by DQN, and a priority experience replay mechanism was introduced to improve learning efficiency. Simulation results show that the method could well solve the problem of overestimation of traditional DQN, accelerate the convergence of neural network, effectively reduce user congestion, and improve network throughput performance.

Abstract:The development of advanced nano-integrated circuit processes has led to a decreasing threshold charge in microelectronic devices, resulting in an increased rate of soft errors caused by single-event effects in digital circuits. To enhance the radiation resistance of standard cells in integrated circuits, this paper proposes a NAND gate structure that is resistant to single-event transients (SETs). In the triple well process, by shorting the substrate and source of each NMOS transistor in the pull-down network, the radiation resistance of the NAND gate was effectively improved, and the hardening of the proposed NAND gate became more effective as the number of inputs increased. Particle incidence simulation experiments were performed by Sentaurus TCAD software in hybrid simulation mode. For the NMOS transistor connected to the output node, the three-dimensional physical model that has been calibrated by the process was used, and the Spice model provided by the manufacturer was adopted for other MOS transistors. Simulation results show that the proposed two-input NAND in 40 nm process could reduce the output voltage fluctuation amplitude in three-input cases at the linear energy transfer (LET) value of incidence particle of 10 MeV·cm²/mg. Besides, the effect of immunity to single particle incidence was achieved in the input mode with N₂ transistor closed. For the hardened three-input NAND gate, the output voltage disturbance could be reduced by up to 85.4% even in the “worst case”. Therefore, the proposed hardening method for NAND gate has a significant effect against SET.

Abstract:To explore the improvement mechanism of welding round steel at soffit on the flexural performance of eccentric concrete-filled steel tube (CFST) members, we established a numerical model of CFST beams reinforced with round steel by using ABAQUS software and verified the model by test results. By analyzing the bending momentdeflection curve, bending momentaxial strain curve, hoop strain curve, restraint index, and neutral axis offset of eccentric CFST members reinforced with round steel, the improvement mechanism of the flexural performance of the eccentric CFST members was revealed. Besides, the influence of the diameter of the round steel and the slenderness ratio of the beams on the flexural performance of eccentric CFST members reinforced with round steel was analyzed. Results show that welding round steel could lower the position of the neutral axis of the section and increase the hoop strain of the steel tube on the compression side. Therefore, the concrete area in compression was increased, and the restraint effect of the steel tube on the compression side on the concrete was enhanced. Furthermore, the flexural bearing capacity and flexural stiffness of the eccentric CFST members were improved, and the larger the diameter of the round steel, the greater the improvement. The ultimate bending moment of the eccentric CFST members decreased with the increase in the axial compression ratio, and the larger the diameter of round steel and the slenderness ratio of beams, the greater the reduction. Welding round steel had a better effect on improving the bending performance of the eccentric member with a large slenderness ratio, and the larger the axial compression ratio, the better the improvement effect.

Abstract:This paper proposes a low-latency intelligent network data transmission scheduling algorithm for real-time network transmission demand scenarios of low latency, stable transmission, and high quality of experience (QoE). The algorithm consists of two parts: data block queuing control strategy and congestion control strategy. The data block queuing control strategy presents a cost-effective model that integrates the creation time and effective time of data blocks, effectively solving the problem of uneven information transmission under transmission time constraint. The congestion control strategy proposes a deep deterministic policy gradient (DDPG) method based on the Gumbel distribution sampling reparameterization with mixed experience prioritization model, which solves the problem that DDPG is not applicable to the congestion control of discrete network action space and significantly improves the quality of network congestion control by adaptively adjusting the sending parameters through learning. Results show that the proposed queuing algorithm could effectively improve QoE in real-time transmission scenarios, and the improved DDPG for congestion control could significantly reduce transmission delay. In the same scenario, compared with traditional network data transmission scheduling algorithms, by integrating the proposed queuing and congestion control strategies, the improved intelligent network data transmission scheduling algorithm could maintain a good balance between low latency and stable transmission and provide higher data transmission quality.

Abstract:To investigate the ground motion intensity measures suitable for evaluating high-rise structures under near-fault ground motions with pulse-like effect, this paper proposes a new ground motion intensity measure considering period elongation effect and higher mode effect based on acceleration spectrum. Taking two high-rise reinforced chimney structures (120 m and 240 m) as research objects, the correlation between damage indices (ParkAng damage index, maximum inter-story drift ratio, maximum structural curvature, maximum floor acceleration, and maximum roof displacement) of high-rise structures and 37 ground motion intensity measures was studied under near-fault ground motions using OpenSEES. Results show that the proposed intensity measure was the optimal index in predicting the ParkAng damage of high-rise concrete structures under near-fault ground motions. High correlation between velocity-related intensity measures and structural damage index was observed. As the structural period increased, the correlation between damage indices and displacement-related intensity measures was improved. Besides, peak ground acceleration had limitations in characterizing the deformation and failure of high-rise structures, but it could be used to analyze the seismic performance of non-structural components. The research results can provide reference for selecting proper measures and structural damage indices to evaluate the seismic performance of high-rise structures under near-fault ground motions.