Abstract:To address the challenge of effectively recognizing unknown modulation types in signal modulation recognition applications using deep learning models, this paper introduces a novel recognition model based on zero-shot learning and autoencoders for open set signal modulation recognition. Features of the modulation signals are extracted through an autoencoder, which incorporates cross-entropy loss, center loss, and reconstruction loss to ensure effective separation of features across different modulation types. Further, open set recognition of modulation signals is conducted based on the distribution of features in the feature space. Additionally, by incorporating the reconstructed signals back into training, the model′s recognition accuracy is significantly enhanced. Experimental results demonstrate that the proposed model not only distinguishes unknown classes effectively, achieving an unknown class recognition rate of 80%, but also maintains a stable known class recognition rate of approximately 95%, outperforming traditional open set recognition methods.

Abstract:The correct identification of partial discharge (PD) as an early indication of many insulation problems in electrical equipment is crucial for formulating maintenance plans which is an effective way to avoid catastrophic failure as the associated defects are treated at an early stage. The memory-based Multimodal multivariant optimization algorithm(MOA) is applied for PD fault identification based on an iterative global and local search to further improve the accuracy of partial discharge (PD) fault identification. However, the construction of the algorithm′s search element has randomness and the setting of related parameters has high pertinence to the identification of complex PD faults in the actual environment. So This article proposes a novel adaptive multivariant optimization algorithm for PD recognition(AMOA).The first step is concerned with the PD data projection into different grids, in which the data may be removed from the data set if it has sparse local density and number of data peak density points are explored as potential PD fault categories if it has high density. After that, the memory-based MOA is applied to identify the PD fault based on an iterative global and local search. With a view to examining the validity of the proposed method, it is applied to the PD datasets of corona discharge, suspension discharge, air gap discharge, discharge along the surface in high-voltage equipment，as well as to the PD datasets of Insulator Surface discharge in GIS under actual operating conditions. The results show that it’s average recognition accuracy is 19.53%、13.04%、19.46%、37.18%、7.79%、8.13%and 4.19% higher than that obtained by the RDB, KPP, SVM-KNN, DPC-DLP, GWOKM, PSO, and MOA algorithms, respectively. It could be concluded that the proposed approach offers the advantages of high PD fault recognition for the electrical equipment.

Abstract:Aiming at the application reliability of the next generation of new electronic nano-devices, the storage-computing 1T1R nano-devices in array actively integrated RNVM with MOSFET based on the 28 nm CMOS process were designed and fabricated, and its comprehensive electrical performances were tested and evaluated in terms of switching ratio(107-8), operating voltage(±1 V), storage windows and so on. The specific reliability experiments were designed and implemented. The results indicated that the unique failure phenomena which did not occur separately in discrete devices truly existed in 1T1R nano-devices in array including the Ion/Ileak degradation (-44.90%/751.64%) of MOSFET in stress and the reverse hard-breakdown of RRAM during cycling tolerance. Taking the microscopic physics mechanism of nano-device into account, the conclusions were summarized that the unique reliability principles triggered by high source-drain voltage and weak gate-control conditions were attributed to the complex micro interaction mechanisms due to its unique structural features and operating modes of 1T1R nano-devices in array. The pertinently specialized test regulation schemes were proposed to improve the reliability of 1T1R nano-devices in array. References for resolving the unique reliability issues caused by the integration of RNVM nanotechnology with logic devices at 28 nm CMOS nodes and below were provided.

Abstract:Wave impact pressure is one of the important physical quantities in ocean engineering. The traditional method of measuring wave impact pressure using pressure sensors has single measurement results and poor stability. Therefore, it is necessary to find a new measuring element to replace pressure sensors. In this paper, MFC is used to measure the wave impact pressure of the model for a wave impact on a vertical plate. In order to verify the feasibility and accuracy of this measurement method, the measured values of wave impact pressure measured by MFC were compared with the calculated values of empirical formulas commonly used in ocean engineering. The experimental results show that the measurement results of MFC are in good agreement with the calculation results of our country′s specifications; compared with our country′s specifications, among all the effective data generated by the five incident wave heights, only the average error of the incident wave height of 5 cm exceeds 10%, which is 18.25%. The average error of the incident wave height of 12 cm is the smallest, with the smallest being 2.56%.

Abstract:In order to improve the accuracy and reliability of rolling bearing fault diagnosis in intelligent manufacturing mode, a fault diagnosis method based on Variational Mode Decomposition (VMD) and time-frequency domain entropy combined with improved Egret Swarm Algorithm (IESOA) to optimize BP neural network was proposed. Firstly, with the help of variational mode decomposition, the problem of pattern aliasing was successfully solved. Secondly, the time-domain Shannon entropy and frequency-domain spectral entropy of each modal component were extracted to construct fault feature vectors as input to the fault diagnosis model. Thirdly, the Halton sequence was introduced to initialize the egret population, the global optimization ability of the egret population optimization algorithm was enhanced, and the improved egret population algorithm was constructed to optimize the BP neural network (IESOA-BP), and finally the bearing dataset of Case Western Reserve University in United States was used for simulation. The results show that the entropy in the frequency domain of VMD time-added is more abundant in the characterization of fault characteristics. Compared with the traditional methods such as BP, PSO-BP, SSA-BP, ESOA-BP and SCESOA-BP, the IESOA-BP method shows higher classification accuracy and better stability in the fault diagnosis of rolling bearings.

Abstract:When applying techniques such as panoramic stitching or video fusion to outdoor environments, complex scenes and lighting conditions often lead to a decline in the algorithm’s keypoint detection capability. Curvature is a stable mathematical feature that describes image edges and exhibits good stability under complex scenes and lighting conditions. This paper delves into the extraction of multi-scale curvature features in image stitching and the Hellinger kernel transformation of the SIFT operator, proposing a multi-scale curvature feature image stitching algorithm based on Shi-Tomasi and RootSIFT. Firstly, the multi-scale Shi-Tomasi method is used to extract illumination-stable keypoints at different resolutions from Gaussian-blurred preprocessed images, making the algorithm more suitable for handling complex environments. Secondly, the RootSIFT enhanced by the Hellinger kernel transformation strengthens the multi-scale feature extraction process, making it more robust to changes in illumination and noise in Euclidean distance. Additionally, FLANN fast matching demonstrates high efficiency in processing large-scale data. Finally, in transformation estimation, the improved PROSAC algorithm of RANSAC can further enhance the speed and quality of stitching. Experimental results on detection performance show that the proposed algorithm can more accurately detect the curvature information of image edges, with feature detection capability improved by 51% compared to the original SIFT algorithm and by 182% compared to single-scale algorithms. The comparative results of multi-scale parameter groups indicate that the algorithm can achieve further optimization, comprehensively enhancing detection capability and real-time performance, demonstrating good adaptability.

Abstract:Accurate prediction of wind power plays an important role in the stable operation of the energy system and power dispatch. Due to the stochastic, intermittent, and nonlinear characteristics of wind power sequences, the use of traditional prediction and a single prediction model often suffers from low prediction accuracy and is easily interfered by noise. In order to improve the accuracy of wind power prediction, a method combining CEEMDAN decomposition technology and neural network model is proposed in this paper. Firstly, the wind power sequence is decomposed into a number of intrinsic mode components by the CEEMDAN method. The complexity of each mode component is calculated by the sample entropy value, and the different intrinsic mode components are reorganized into reconstructed subsequences based on the sample entropy values. Middle and high-frequency sequence data are predicted using the BiLSTM model, while middle and low-frequency sequence data are predicted using the TCN model. Finally, the predicted values from the different models are combined to obtain the final prediction. Through simulation experiments, the results demonstrate that the model proposed in this paper achieves the lowest values in the evaluation metrics RMSE, MAE, and SMAPE, and the highest value in the R-squared metric. The average values of these indicators are 91.413 2 MW, 53.517 3 MW, 22.263 8 MW, and 0.980 7, respectively, which are better than those of the comparison models. This indicates that the model presented in this paper has high accuracy.

Abstract:To address the issues of high fingerprint collection costs and the difficulty of constructing datasets in indoor positioning, a method for indoor positioning fingerprint generation based on a conditional sparse autoencoder generative adversarial network is proposed. This method enhances the feature extraction capability by adding hidden and output layers to the autoencoder, guiding the generator to learn and generate key features of fingerprint data. A fingerprint selection algorithm is used to filter out the most relevant fingerprint data, which is then added to the fingerprint database and used to train a convolutional long shortterm memory network model for online performance evaluation. Experimental results show that the conditional sparse autoencoder generative adversarial network improves the accuracy of indoor positioning in multi-building, multi-floor environments without increasing the number of collected samples. Compared to the original conditional generative adversarial network model, the positioning error in predictions on the UJIIndoorLoc dataset is reduced by 6%, and in practical applications, the positioning error is reduced by 14%.

Abstract:FM continuous terahertz wave system is limited by the excitation device, resulting in a generally narrow bandwidth, in the detection of samples due to the limited amount of data at a single point, there is not easy to extract the feature signal, to address this problem, this paper proposes a chirp-z transform algorithm based on the Kaiser window. First of all, the original signal is processed by adding windows, and the linear frequency modulated z-transform is used to extract the feature signals of the depth defects in different samples, and different signal processing methods are introduced for comparative analysis, and the results show that the method proposed in this paper reduces the computational complexity while retaining the feature information of the samples in the limited amount of data, comparing with different kinds of composite material detection data processing, which proves the validity and universality of the method.

Abstract:In order to effectively utilize the regenerative energy returned by the elevator and reduce the energy loss of the system, this paper focuses on the energy storage characteristics of the ultracapacitor, Firstly, the structure and principle of elevator energy storage system based on supercapacitor are introduced; secondly, the small signal model of bidirectional DC/DC converter is established, and the double closed-loop control structure and parameter design process of bidirectional DC/DC converter based on traditional PI controller are discussed; then, on the basis of analyzing the principle of ant-lion optimization algorithm, a PI parameter design method based on ALO algorithm is proposed. Finally, the simulation model of elevator energy storage system based on ULtracapacitor is established by using MATLAB simulation software. The simulation results show that the ultracapacitor can timely and accurately recover the braking energy and return the recovered energy to the DC bus, ensuring the voltage stability of the DC bus and reducing the energy consumption of the system, which fully verifies the effectiveness and superiority of the method.

Abstract:To address the challenge of ensuring safety, efficiency, and smoothness in collision avoidance decisions for unmanned surface vessel in complex environments with dynamically changing obstacles, we propose a collision avoidance method that combines the dynamic window approach and DDPG algorithm. Firstly, in the traditional collision risk model, the distance to closest point of approach and the bearing angle at the closest point of encounter are added as evaluation factors to make the risk evaluation of the unmanned surface vessel more reasonable. Next, we design a local guidance method based on dynamic window approach, the reachable position of unmanned surface vessel by dynamic window approach is used as the local guide point, and the guide reward is added to increase the action reward near the guide point, so that DDPG algorithm can obtain more accurate updating direction in training. Finally, the method is tested in various obstacle environments. Experimental results show that compared to the traditional DDPG algorithm, the proposed method generates more reasonable, smoother and less risky paths. Additionally, it improves convergence speed by approximately 37.5%, verifying the effectiveness of the proposed method.

Abstract:As a key component of autonomous driving technology, object detection technology is crucial for vehicles to achieve autonomous navigation and decision-making functions. Existing algorithms still face difficulties in meeting both detection accuracy and detection speed simultaneously. In this regard, a real-time object detection algorithm FSSD-DETR based on RT-DETR is proposed. This algorithm introduces FADC module into the backbone to optimize the feature extraction process. A small object detection layer is introduced to improve the detection performance of small targets for distant vehicles and pedestrians. Based on the SSFF module and TFE module, the neck network has been redesigned to improve the accuracy of detection. The DySample upsampling operator is used to replace Nearest Neighbor Interpolation to improve possible issues such as detail loss, jagged edges and image distortion. The experimental results show that compared to the original RT-DETR model, the improved algorithm has increased mAP by 3.6% and 2.1% on the SODA10M and BDD100K datasets respectively.The experiment demonstrates that FSSD-DETR significantly improves the detection accuracy while ensuring real-time performance, which has application value.

Abstract:To address issues such as confusion in distinguishing left-right and front-back categories due to similar features in current vehicle orientation scene recognition tasks, we proposed a vehicle orientation scene recognition method that integrates global-local attention. We introduced the concept of multi-view vehicle scenes, utilized OSMNet for feature extraction and scene classification, and developed a global-local attention module to focus on key areas across different orientation scenes for effective spatial orientation learning. Additionally, we designed a global-local positional attention module to address overlapping class distances between certain vehicle orientation scenes.Experiments on an 8-class scene dataset demonstrated that our D-CBAM and HGLP modules effectively enhanced the capture of global and local information in feature maps, improving model recognition accuracy by 3.54% and 4.22%, respectively, in ablation studies. Comparative experiments showed that our model achieved an accuracy of 95.49%, which is 5.46% higher than the baseline model. Overall, our model outperformed other classification models in recognizing most orientations better than the baseline model. These results demonstrate that our improved classification model effectively learns vehicle orientation information, bridging the gap for matching images from distant, intermediate, and near perspectives, and laying a foundation for tasks such as multi-part vehicle detection and segmentation.

Abstract:Metaphorical affective prediction can help improve the user experience of social media content, while also having potential value in mental health monitoring and virtual psychotherapy. In addition, it can more accurately identify the affective needs of the target audience, optimize advertising strategies, and improve business efficiency. In order to further enhance the effectiveness of metaphorical affective prediction, architecture on multi-mode metaphorical affective prediction method that consolidating intra-class difference and inter-class coherence is proposed. Firstly, three single-mode models are introduced, including image semantic model, text semantic model, and voice semantic model, to extract personalized differential features from three data sources, respectively. Then, a deep layering multi-mode model is introduced to learn the coherences between multiple modes through intermediate layer fusion, better utilizing the complementary information provided by bi-modal and tri-modal data. Finally, the four aforementioned models are fused using a decision-making layer fusion approach to predict multi-modal metaphorical feelings in an end-to-end architecture. Extensive ablation experiments and comparative studies conducted on open-source datasets have demonstrated the effectiveness of proposed approach.

Abstract:Aiming at the problems of irregular morphology and suboptimal detection performance of surface defect on aluminum ingot alloys, an improved YOLOv5-based defect detection method is proposed. Firstly, The Res2Net feature extraction network block is employed to replace the CSPDarknet53 module of the baseline model, which can effectively detect the multi-scale defect. Secondly, the CBAM convolutional attention module is introduced into the backbone network of YOLOv5 to enhance the representational ability of defect features. Finally, the over-parameterized reparameterization convolutional blocks are used to substitute for the 3×3 convolutional blocks in the backbone and neck networks so as to reduce the model′s inference latency. Experimental results compared with the traditional target detection methods demonstrate the improved method achieves a mAP of 75.8% for defect detection, which is a significant improvement both in detection accuracy and inference speed, and can well satisfy the tasks and demands of practical industrial production.

Abstract:Aiming at the problems of omission, misdetection and low detection accuracy of pedestrian crossing guardrail detection in the complex scenarios of occlusion, dense multi-target situations as well as multiple people overtopping, a multi-scale feature fusion and enhancement algorithm for pedestrian overtopping guardrail detection is proposed. Firstly, an algorithm based on Dual Vision Transformer and SCConv, which is applied to the backbone network, enhances the capture of global context information and finer-grained information, and improves the local fine feature extraction and feature fusion capability of the network; second, a multi-scale feature fusion enhancement module AM-SPPFCSPC is proposed, which compensates for the feature loss caused by maximal pooling, improves the feature map. The richness and completeness of the feature map is improved, and the multiscale feature extraction and feature fusion capability is enhanced; in addition, the feature fusion layer is further refined by replacing the ordinary convolution with GSConv and designing the VOV-GSCCSP module based on GSConv and SCConv, which effectively reduces the computational cost and the complexity of the model, while maintaining a higher degree of accuracy; finally, a highly efficient multi-scale feature fusion module, AM-SPPFCSPC, is introduced in the trunk to reduce the complex background and the complex background and the complex background and the complex background. Attention EMA, which reduces the interference of irrelevant targets in the complex background and fuses the multiscale information to achieve richer feature aggregation. The experimental results on the homemade pedestrian over guardrail dataset show that the proposed algorithm in this paper achieves 93.6% mAP with the addition of fewer parameters, which is 4.5% higher than that of the original model, and has a detection speed of 108.5 FPS, which improves the problems of leakage, false detection and low detection accuracy, while still having a high real-time performance, and is more suitable for real-time detection of pedestrians crossing the guardrail.

Abstract:To address the issue of insufficient accuracy in defect segmentation caused by weak expression of surface defect characteristics on condenser copper tubes and feature confusion between similar defects, a feature-optimized method for surface defect segmentation on condenser copper tubes is proposed. Firstly, to address the problem of indistinct surface defects on condenser copper pipes, the method utilizes an attention optimization module based on the defect area attention enhancement strategy to enhance the feature expression ability of defects and suppress background feature expression. Secondly, through the use of dilation convolutions with varying rates and the integration of feature map optimization technology, cross-domain semantic capture of pixels is achieved and resolve the issue of feature confusion between similar defects. Finally, a multi-scale feature enhancement fusion method based on feature alignment is established to improve the model′s detection ability for defects at different scales. Multiple sets of comparative experiments are conducted on images of condenser copper tubes which are captured in real production line environments, and the results show that the proposed method achieves the balance between the precision and the number of parameters when solving the above problems, and achieves a good segmentation effect. The algorithm achieves an average intersection over union of 80.53% and a Dice coefficient of 88.94%, with the model size being only 25 MB.

Abstract:In order to solve the problem of complex and diverse warehouse environment and the low performance of traditional warehouse object detection models, this paper proposes an improved YOLOv5 (You Only Look Once version 5) warehouse object detection model YOLOv5-CE (YOLOv5-ConvNeXt EIoU) which based on the PaddlePaddle framework. Firstly, to improve the detection of warehouse objects in complex and diverse environments, the ConvNeXt network is used to replace the original YOLOv5 backbone network to improve the feature extraction ability of small and medium-sized warehouse objects. Secondly, in order to improve the convergence speed of the model and the detection accuracy of objects, EIoU Loss (efficient intersection over union loss) is used to replace the loss function of the original model. Finally, by using the self-made warehousing training set to carry out multi-model comparison experiments. The experimental results show that when detecting cargo, tray and forklift, the average detection accuracy of the improved model (mAP@0.5:0.95, mean average precision@0.5:0.95) reaches 89.8%, which is 1.1 percentage points higher than the original YOLOv5, of which 4.2 percentage points is increased in small-scale warehousing objects; in the detection of medium and large-scale warehouse objects, it increased by 1 percentage point. The average recall rate for small warehouse objects increased from 61.1% to 66.8%. Compared with other models such as YOLOv6, YOLOX, YOLOv7, and Faster R-CNN, YOLOv5-CE all shows better accuracy. At the same time, in view of the above model, YOLOv5-CE also achieves a good balance in the number of model parameters, detection speed and detection accuracy, which can better meet the precise detection of warehouse objects of different sizes and types.

Abstract:Aiming at the current challenges of pedestrian detection, such as complex environments, variable target sizes, and severe occlusions, which cause existing detection techniques to be prone to misjudgment and omission when recognising dense pedestrians, this paper proposes an efficient YOLOv8 improved model for dense pedestrian detection in complex scenes. DCNv2 is introduced into the backbone network, and C2f_DCNetv2 is designed to replace the C2f module, which improves the feature extraction capability of the backbone network; the detection capability of the model for small targets is improved by adding small-target detecting heads to the architecture, which improves the accuracy of small-target detection and recognition; based on the four detecting heads as well as the AFPN, the AFPN-4H is designed, which optimises the information fusion between the feature layers and improves the model′s adaptability and detection accuracy for targets of different scales; finally, through the combination of Wise-IoU, Focaler-IoU, and MPDIoU, the WFM-IoU is obtained, which further improves the target localisation accuracy. The experimental results show that compared with the original YOLOv8n model, it improves 1.6, 4.0, 3.6 and 3.8 percentage points in the key indexes of P, R, AP50 and AP50:95, respectively, which are also inferior to other algorithms. The improved algorithm in this paper has better performance in the dense pedestrian detection task in complex scenes.

Abstract:Addressing the issue of erroneous extraction of fine structures such as human hair in image matting tasks, the problem essentially stemmed from inaccurate prediction of pixel alpha mattes due to mixed information within these regions. To address this problem, a novel end-to-end hierarchical feature aggregation matting network model is proposed. This model incorporates a shared encoder and two independent decoders, leveraging channel and positional attention mechanisms to aggregate low-level texture clues and high-level semantic information in a hierarchical manner. It enables perceiving foreground transparency masks from fine boundaries of individual portraits and adaptive semantics without additional inputs. To guide the hierarchical feature aggregation matting network model in refining the overall foreground structure and restoring hair texture details, cross-entropy loss, alpha matte prediction loss for unknown regions, and structural losses are integrated. To validate the effectiveness of the proposed model, experiments were conducted on the self-constructed MCP-1k dataset and the publicly available P3M-500-NP dataset. Experimental results demonstrated that the proposed model achieved errors of 0.0076 MSE and 25.59 SAD on MCP-1k dataset, and 0.0072 MSE and 25.52 SAD on P3M-500-NP dataset, respectively. Compared with other typical deep matting models, it showed significant improvements in restoring fine human hair and enhancing semantic structure in portraits, effectively addressing the issue of erroneous extraction in human hair regions.

Abstract:To solve the problems of low efficiency and long time-consuming in the recovery of unexploded ordnance underwater by LBL positioning system, and the problem of insufficient positioning accuracy by USBL system, an efficient single beacon underwater positioning system is designed. This system consists of two parts: one part is an ultrasonic beacon, which is launched into the water simultaneously with the experimental underwater shell; the other part is an ultrasonic receiving module, which is mounted on the UUV to detect the ultrasonic signals from the beacon. By capturing the ultrasonic signals emitted by the beacon multiple times along its pre-determined navigation path, and recording the corresponding time stamp data, the system can calculate the precise three-dimensional coordinates of the unexploded projectile using VLBL positioning technology. After a series of experimental tests, it has been proven that this single-beacon underwater positioning system can locate unexploded experimental projectiles within a range of 50 meters in lake environments, with a positioning accuracy that can reach the decimeter level. The successful development of this technology not only improves the detection efficiency and safety of underwater unexploded ordnance but also provides a new high-precision positioning method for other marine engineering operations, underwater archaeology, and seabed resource exploration.

Abstract:Regarding the problem of insufficient identification accuracy caused by noise interference and difficult determination of key parameters when using the variational mode decomposition (VMD) algorithm to decompose the sub-synchronous oscillation signals generated during the grid connection process of wind power, this paper proposes a signal decomposition algorithm based on wavelet threshold denoising (WTD) and genetic algorithm (GA) optimized VMD, combining with the sub-synchronous oscillation mode identification method of autoregressive moving average model (ARMA). Firstly, wavelet threshold denoising is used to process the active power output of the wind turbine; secondly, VMD is used to decompose the denoised signal, obtaining K intrinsic mode components. In order to achieve the optimal VMD decomposition effect, an adaptive genetic algorithm is used to optimize the penalty factor α and the number of decomposition layers K. Finally, the signal is restructured and an ARMA model is established to directly identify the frequency and damping ratio of the sub-synchronous oscillation signal. By building a simulation experiment platform for direct-drive wind turbine grid connection model and collecting sub-synchronous oscillation signals for mode identification, the simulation results show that, compared with other identification algorithms, the proposed VMD-based method has better feasibility and superiority.