Patches, as a valuable piece of information for security-related tasks, are often missing in security advisories. In this article, we propose an automated approach, named PatFinder, to find and parse patches for open source software (OSS) vulnerabilities. First, PatFinder identified commits from numerous vulnerability-related references. Then, PatFinder selected patches based on code changes of identified commits and a weighted voting mechanism. Finally, based on designed patch parsing methods, metadata of patches (i.e., paths of modified files and names of functions) was obtained. Our experiment has shown that PatFinder can achieve a coverage of 73.10% and a recall of 0.802, significantly improving the coverage and recall of existing approaches.

In order to store, share, apply and mine more valuable information for residents' health, this paper designs a Hadoop-based regional health big data platform. The platform stored health data from various health institutions and physical examination institutions in the pre-database through data collection. The medical data was stored in the standard library through data governance and data exchange, and data was opened out in the form of data services to provide health big data support to medical institutions, physical examination institutions, health management companies, etc., which promoted the development of medical health products research and development, and better provided residents with medical and health services, and ultimately achieved full life cycle health data management.

In order to solve the problem that it is difficult to explain the prediction results of just-in-time software defects, based on the improved model of polynomial neural network, an ante-hoc interpretable just-in-time software defect prediction method is proposed. This method formalized the causal relationship between code metric elements and prediction results, and outputted it as a K-G polynomial function. The standardized regression coefficient was used to measure the importance of metric elements to analyze the causes of the defects. The experimental results show that on the premise that the average prediction accuracy reaches 0.797, it has good interpretability at the same time.

Prostate cancer ranks as the second most frequently diagnosed neoplasia and the fifth leading cause of mortality in male patients with cancer. It is of great clinical significance to design an image-assisted diagnosis system for prostate pathological section. In the case of only image-level annotation data set, convolutional neural network (CNN) is used to classify only images, but no cancerous regions are given. In view of this situation, the CNN model with efficientnet-B0 architecture was used as the basic classification model, the image was divided into patches and the categories of each patch were obtained, and the cancerous regions were obtained by clustering algorithm. Pathological images were uploaded on the Web front end, and auxiliary diagnosis results were viewed after the processing was completed. Experimental results show that the precision of the system is 76.61%, and the recall rate is 72.52%, which can effectively obtain the general area and obtain satisfactory auxiliary diagnosis effect.

The core task of places such as power sales offices is to recognize the user's intention, and current intention recognition methods require a large amount of data to assist in model training. But for these places, it is very difficult to collect data on a large scale. Therefore, it is very important to utilize the training samples efficiently based on the limited number of samples in the dataset. In summary, this paper proposes a semantic distance-based curriculum learning strategy for the task of electric power intent recognition, which can train and learn the samples more efficiently. The experimental results show that the curriculum learning strategy can significantly improve the recognition accuracy of the business on the task of electricity business hall intention recognition.

The 2D NMR core analysis experiments play an important role in unconventional oil and gas exploration and development, but the sample experimental equipment only provides raw echo data and rough 2D spectra, which require secondary processing of the results to obtain oil and gas saturation and other parameters. A set of general-purpose 2D NMR experimental data analysis softwareis developed. Based on the fast high-resolution 2D spectral inversion algorithm and the area integration algorithm, fine processing and analysis of 2D NMR experimental data were achieved, and quantitative parameters such as porosity and saturation related to oil and gas reservoirs were directly obtained, extending the application of 2D NMR experimental results.

In response to the collaborative and operational demands of healthcare data ecosystems, this study investigated the fundamental models for medical data assetization and governance. We systematically explored methodologies for data assetization and operational frameworks, and proposed a unified model architecture. Aligned with clinical practice requirements, a disease-specific data asset management platform was developed. Platform implementation and effectiveness analysis demonstrate that our model significantly enhances cross-institutional collaboration and value-driven utilization of specialized medical data, providing critical insights for optimizing healthcare data interoperability and maximizing its socioeconomic impact.

The water supply network simulation system has a complex topology and dense component data. The efficient simulation of the system puts forward high requirements on the performance of the implementation platform. The "Songshan" supercomputer system adopted a CPU + DCU heterogeneous architecture, in order to realize the efficient operation of simulation calculation on this architecture. According to the HIP heterogeneous programming model, the computing hotspots were mapped to the DCU accelerator to achieve, and the MULTPARA multi-level parallel optimization method was used to achieve the concurrency of different component calculations on the DCU accelerator, to solve the problem that the thread blocks could not be fully parallelized. The cross-thread access was used to solve the thread divergence problem existing in some component calculations. The experimental results show that compared with the previous implementation platform, it has achieved a speedup of 8.537.

To improve the trajectory tracking accuracy of coaxial multi-rotor UAV under airflow disturbance and rotor fault, a fast terminal sliding-mode fault-tolerant control method is proposed. The fault model of a coaxial 12-rotor UAV was established. A fast terminal sliding-mode control law with a motion loop was designed to convert the trajectory command into the velocity command, and the air flow disturbance and rotor fault were estimated by using the slidingmode adaptive law and the sliding-mode observer respectively. A fast terminal sliding-mode fault-tolerant control law was designed for the speed loop to achieve high precision of the coaxial 12-rotor UAV trajectory tracking. The simulation results show that the proposed method can compensate for the rotor fault and airflow disturbance more effectively, compared with the integrated sliding-mode fault-tolerant control method, and the maximum trajectory tracking error is only 0.3 m, which can accurately estimate the rotor fault within 0.2 s, and the maximum air disturbance estimation error is only 0.2 m/s2, which has great trajectory tracking effect.

In order to accurately and efficiently identify and locate the bird droppings on small photovoltaic power station, the improved YOLOv5 model is carried on the Raspberry Pi to form a bird droppings detection system of photovoltaic power plants. The system reduced the threshold of confidence to identify all suspicious bird droppings, identified and partitioned single photovoltaic panels, and increased the confidence threshold to accurately detect suspicious bird droppings in photovoltaic panels. In order to make the YOLOv5 algorithm more suitable for detection, the pyramid split attention was integrated in the algorithm. The small target detection layer was added and the original pooling operation was replaced by SoftPool. In the test set, the mAP~~0.5 of PV-YOLOv5 model identified for photovoltaic panels was 96.78%, which was 2.35 percentage points higher than that of Faster-RCNN. The mAP~~0.5 of NF-YOLOv5 for bird droppings recognition was 94.12%, which was 5.8 percentage points higher than the original YOLOv5 model.

The width of the tobacco fiber is an important indicator of the quality of the cigarette. In order to be able to control the quality of tobacco fiber width in real-time, this paper designs a real-time online detection method of tobacco fiber width. The SPP and PANet structure were fused in YOLOv4, and the three output scales were changed into one output scale making the improved YOLOV4 algorithm simpler, and the calculation speed increased nearly 35% in the same case of performance. The width calculation method was the MCS (Moving Center Search) center mobile search method, which could automatically search for both sides of the tobacco fiber, and more accurately calculate the width with the calculation accuracy of 0.2 mm. The MCS tobacco fiber width calculation method was combined with the improved YOLOv4 model of this paper, and real-time online tobacco fiber width detection was performed to control the quality of tobacco fiber in real-time and improve production efficiency.

Aimed at the problems of dynamic environment and unpredictable scene in unmanned aerial vehicle (UAV) autonomous search mission planning, the classical simple hierarchical ordered planner (SHOP) method is improved, and an iterative mission planning model with multiple planning and multiple execution is proposed based on the improved SHOP method. It provides a feasible solution to the mission planning problem with incomplete information. The simulation environment was built, taking the autonomous search task of UAV as an example, the feasibility and practicability of the iterative mission planning model were verified. Compared with the SHOP planning method, the advantages of the iterative planning method are illustrated.

A method for calculating the importance of nodes in electronic information wireless communication network based on efficiency matrix is proposed. The local information and global information of the node in the electronic information wireless communication network were obtained through the transmission efficiency matrix, the triangular fuzzy number method was used to construct the importance evaluation matrix according to the obtained information, and the importance evaluation matrix was used to calculate the importance of the node in the electronic information wireless communication network. The experimental results show that the importance measurement results of electronic information wireless communication network nodes by the proposed method are consistent with the actual test results, which improves the measurement accuracy and shortens the running time.

As underwater acoustic channel is an open environment, underwater wireless sensor network (USWN) is subjected to other attacks and threats. To solve the security problem of UWSN identity authentication, we propose a UEIAP (underwater ECC identity authentication protocol). UEIAP implemented three-factor bidirectional identity authentication, which made anonymous identity dynamic and negotiates temporary session keys for secure communication in the future. Strict BAN logic formal proof and AVISPA simulation experiment show that the protocol meets the security and functional requirements with relatively little overhead and is suitable for UWSN.

In two-way full-duplex relay systems based on non-orthogonal multiple access (NOMA) for Beyond the Fifth Generation (B5G) networks, we propose a power control mechanism using successive convex approximation (SCA) to coordinate co-channel interference and improve spectral efficiency, which efficiently obtains suboptimal solutions for the original non-convex problem within polynomial time. To satisfy the low-latency requirements of B5G mobile networks, we further design a deep neural network (DNN) -based online power control mechanism. Simulations verify the effectiveness of both mechanisms, showing that the DNN-based approach achieves comparable performance to the SCA-based method while significantly reducing online computational time.

In order to solve the problem of high complexity and many influencing factors in existing energy-efficiency optimization algorithms in IRS-assisted millimeter wave communication systems, we propose a high energy-efficiency optimization algorithm based on fractional programming theory. Specifically, we used Dinkelbach and successive convex approximation techniques to jointly optimize the hybrid precoding at the base station and the phase shift matrix of IRS. The simulation results show that, compared with the traditional system without intelligent reflecting surface, the energy efficiency of the system under the joint design of hybrid precoding and phase shift matrix is improved by about 215%. The results also show that deploying more low-cost passive elements on reflecting surface can significantly improve the downlink energy efficiency of the system.

During the hardware implementation of the high efficiency video coding standard (HEVC), its high computational complexity and high data dependence not only hinder the performance of real-time video encoding, but also bring higher resource consumption. In this paper, we analyzed the data/timing dependency relationship between the various algorithms of the intra-frame loop when processing pixel blocks, and the HEVC intra-loop reconfigurable array structure based on the variable coding block pipeline is completed. By using the adjacent interconnection interface between the modules, the handshake communication realized the pipeline processing between the entire intra-frame loop coding blocks, and improved the calculation efficiency of the intra-frame loop. The experimental results show that the proposed method meets the requirements of computing speed and hardware resources. Compared with the serial-based intra-loop encoding processing circuit, the execution time is reduced by 87%, and the clock frequency reaches 125 MHz.

Aiming at the low efficiency of keyframe extraction in content-based video retrieval, resulting in insufficient representation of selected keyframes and performance of the entire video retrieval system, this paper proposes a keyframe extraction algorithm based on multi-feature fusion similarity. A combination method of color histogram and full convolutional neural network was used to detect video shots, and segmented the video into shots with higher content correlation. The multi-feature fusion similarity method was used to extract keyframes from the segmented shots. This paper used the deep feature similarity method to remove redundant keyframes, and obtained more accurate results. Experimental results shows that the extracted keyframes have a strong generality for video, and can be applied to video retrieval and summary. The overall recall and precision rate can reach 85.61% and 83.21%, respectively. Compared with other algorithms, the redundancy of the key frames extracted by this algorithm is relatively small.

Recently, syntax-aware neural semantic role labeling (SRL) has received much attention. However, most of previous syntax-aware SRL works exploit homogeneous syntactic knowledge from a single syntactic treebank. Considering several high-quality publicly available Chinese syntactic treebanks, this paper proposes to extend graph convolutional networks (GCNs) for encoding heterogeneous syntactic knowledge in the heterogeneous dependency trees and makes a through comparison on various encoding methods to improve SRL performance. This model achieved 84.16 and 85.30 F1 on CPB 1.0 and CONLL-2009 Chinese data sets, respectively, both outperforming the corresponding homogeneous syntax-aware SRL models and significantly improving the performance of previous methods.

For high-voltage transmission lines, there will be foreign bodies affecting the normal operation of the line. This paper proposes an improved CenterNet method for detecting foreign bodies in high-voltage transmission line inspection to detect common small targets of foreign bodies in transmission lines. In this method, a wide residual module was designed based on cavity convolution, and the feature extraction surface of the residual module was widened. The channel attention mechanism and spatial attention mechanism were used to process the feature information twice, and the detection accuracy was optimized. The experimental results show that the improved average accuracy is improved from 88.59% to 93.09%, and the overall accuracy is improved by 4.5 percentage points.

Aimed at the problem that the CenterNet detection algorithm has a large number of network parameters and fails to fully and effectively utilize the multi-scale local region features, an MIR-SPPA-CenterNet detection method is proposed to improve the CenterNet detection network. Specifically, mixed invert residual (MIR) block was introduced into the backbone network of CenterNet to achieve a lightweight effect. In addition, an improved spatial pyramid pooling with attention (SPPA) block was introduced to pool, cascade, and filter multi-scale local area features so that the network could adaptively learn more comprehensive and effective target features. Experiments show that this method has better detection results on the general PASCAL VOC dataset and the self-built L-KITTI dataset.

At present, the number of patients withchronic insomnia disorder (CID) is increasing year by year. Timely diagnosis can effectively avoid the aggravation of symptoms of CID patients. Magnetic resonance imaging (MRI) technology combined with a classification algorithm can be used to identify CID patients. The traditional MRI data classification algorithm is based on single-mode feature SVM algorithm, but this algorithm has poor effect on CID patient data classification. Therefore, a CID patient recognition algorithm based on multimodal feature integration is proposed to achieve better results. The multimodal feature integration algorithm mapped multimodal features based on resting-state functional MRI technology and used the integration algorithm for classification and comparison experiments. The experimental results show that, compared with the traditional MRI classification algorithm, the multimodal feature integration algorithm has better classification effect on CID patient data, and can effectively identify CID patients, to carry out relevant medical auxiliary diagnosis.

Gait recognition has a wide range of applications in real life. The key of gait recognition is to extract gait related features from the video frames of walking people. Aimed at the problem that the existing methods can not obtain gait features based on appearance features, using the disentangled representation learning method, an autoencoder architecture was proposed to decompose gait features and appearance features, and the joint entropy based on Renyi entropy was used to minimize the mutual information between gait features and appearance features. Through a large number of experiments on CASIA-B and FVG data sets, this method shows better decoupling ability and higher recognition accuracy in gait recognition.

Aimed at the problems of poor feature fusion ability, weak correlation of time-series information and unclear behavior boundary in the existing human behavior detection methods, a human behavior detection method based on spatio-temporal interactive network is proposed. The dual flow feature extraction module was redesigned, and a connection layer was added between the two networks of spatial flow and spatio-temporal flow. The improved spatial transformation network and visual attention model were introduced into spatial flow and temporal flow networks respectively. A feature fusion module based on pixel filter was designed to calculate the correlation of time series information in key areas and aggregate two kinds of features with different dimensions. The loss function of the network was optimized. Experimental results on AVA dataset show that this method has advantages on detection accuracy, speed and generalization ability.

The pedestrian is easy to be blocked and the scale is different, so the pedestrian missed detection rate is high. In view of this, the pedestrian detection algorithm based on Anchor-free idea is improved. Aimed at the problem that convolutional neural network was sensitive to target scale changes when extracting features, a fine-grained feature fusion strategy was proposed to obtain rich pedestrian feature information. The spatial attention mechanism was used to study the weight of different regions of the feature map to improve the expression ability of the model. Using multi-scale detection method, the model adaptively detected pedestrians of different scales and enhanced the robustness of model detection. The experimental results show that MR－2 of 11.33%, 6.81%, 11.52% and 50.09% are obtained on Reasonable, Bare, Partial and Heavy subsets of Cityperson dataset, respectively, which is better than other pedestrian detection algorithms.

In order to solve the problems of late time and high work intensity in the detection of eggs without sperm, a VGG16 network model was improved and a graphical user interface was developed for hatching 2.5 d eggs. The image of hatching 2.5 d eggs was collected by a self-made static image acquisition device. The improved model achieved 98.82% discrimination accuracy and 97.23% recall rate on the enhanced test set, and the detection time of single image was 97.56 ms. Compared with the original network, the recognition accuracy was improved by 5.56 percentage points, and the recognition time of single image was saved by 14.78 ms. The results show that the improved model can effectively realize the nondestructive identification of egg fertilization information in the early stage of incubation, which provides technical support for the subsequent development of online nondestructive testing devices.

Aimed at the problems of pig skeleton extraction, such as difficulty, low accuracy and long-time consumption, a pig skeleton extraction method based on encoder-decoder network is proposed. The key point heat map generation model was constructed, ResNet50 residual network and U-Net semantic segmentation network were combined to build an encoder-decoder network structure, and the attention mechanism was introduced to improve the feature extraction accuracy of the key points of small targets such as tail and hoof. The offset of key points was predicted while generating the key point heat map, which made up for the accuracy loss when calculating the original position of the key points. The Hough voting mechanism was used to weighted aggregate the two points, and the pig skeleton was finally mapped. The experimental results show that the skeleton extraction accuracy is 85.27%. Compared with the ResNet50 residual network, the accuracy is increased by 22.67 percentage points with similar time consumption. This study provides a new method for pig skeleton extraction, which can provide a technical reference for further pig behavior research.

There are different challenges between natural language and structured knowledge in the field of question answering based on knowledge graphs. Therefore, this paper proposes a method of using predicates to select paths (PMKBQA). A multimodal domain knowledge graph and question set were constructed. Starting from the subject entity identified in the question, calculating the similarity between its edge and the question predicate, and generating the answer path hop by hop until the answer to the question was found. Relevant images of the answer of the question according to the answer path were acquired. The user satisfaction evaluation experiment was done on the domain question set, and the results show that this paper can provide users with satisfactory images. Meanwhile, a question answering effect evaluation experiment is done on the QALD data set, and the results show that our method is better than the baseline method in F1.

The string segmentation task of photovoltaic power station requires to accurately identify the edge of PV string, so as to obtain accurate position information. For this task, this paper designs the scSE-Unet8 semantic segmentation model. The squeeze excitation module (SCSE) was introduced into the U-Net and the complexity of the model was reduced. The model was trained and verified on the aerial image data set of PV power station. The experimental results show that compared with the U-Net, scSE module revise the space and channel features, so as to emphasize the important edge feature information. Therefore, scSE-Unet8 has better effect on string edge pixel detection. After cross validation, the mIoU (Mean Intersection over Union) on the test set is 98.62%. The boundary information extraction algorithm was used to process the output result of scSE-Unet8, eliminate a small amount of false detection and missed detection in the original segmentation result, and the string boundary at the pixel level can be obtained.

Aimed at the problem that the existing image caption models lack attention to the local details of an image and tend to give general description, a Chinese image caption method combining encoder and visual keyword search is proposed. A fusion encoder was constructed, and the local and global features of an image were extracted simultaneously in a convolutional neural network (CNN) to enrich the semantic information of image features in long short-term memory (LSTM) decoding stage. Aimed at the problem of general expression, the image retrieval method based on convolutional neural network was used to find the potential visual words, and was integrated into the word vector generation process in the decoding stage. Reinforcement learning mechanism was introduced to optimize the CIDEr evaluation index at the sentence level to improve the lexical diversity of image description. Experimental results verify the effectiveness of the proposed method.

To promote the development of autonomous driving technology, this study addresses the poor detection performance and low accuracy of existing vehicle detection algorithms for small-sized targets by proposing QF-YOLOv5, an improved YOLOv5-based vehicle detection algorithm. Building upon the YOLOv5 architecture, the following enhancements are introduced: An additional small-scale feature fusion detection layer is incorporated to enhance the detection capability for small targets. An attention mechanism is integrated to guide the network to focus on effective features while suppressing irrelevant ones, thereby improving detection performance. Depthwise separable convolution is adopted to reduce computational complexity. The Mini Batch K-Means clustering algorithm is employed to accelerate network convergence. The Quality Focal loss function is utilized to enable supervised learning for continuous numerical predictions. Experimental results demonstrate that the proposed algorithm achieves improvements in both detection accuracy and real-time performance.

The text-independent speaker verification system is less effective when the test utterance is shorter. In view of this, a method of enhancing acoustic features is proposed to assist the system. The method used a generation model based on seq2seq to generate longer acoustic features from short-term acoustic features. The generation model included an encoder for extracting deep features and a decoder for outputting acoustic features. It used an attention mechanism to obtain the relationship between sequences and added cosine distance loss to improve the generalization performance of the generation model during training. The trained text-independent speaker verification model was used as a component of the generation model training architecture to help the generation model training. The experimental results show that under the condition of 1-3 seconds of speech duration, the equal error rate of the system is reduced by 7.78% on average after using this model.

Due to the anisotropic distribution of ground targets in remote sensing images, such as large variation of aspect ratio and wide range of target scale, the existing segmentation methods are still insufficient for the segmentation ability of targets with long-range banded structure and dense discrete distribution objects. In order to solve this problem, a remote sensing image segmentation network integrating anisotropic context is proposed. The network extracted the gradient information of the target by imposing a priori constraints on the gradient convolution kernel parameters, optimized the segmentation edge. It designed modules such as multiscale parallel dilated convolution and anisotropic target composite strip pooling module, and captured the anisotropic context information of different scale targets in remote sensing images. The multi-scale context information was fused and the image details were restored. Experiments on the public Potsdam and Vaihingen datasets show that the anisotropic context fusion network in this paper is superior to the advanced segmentation networks such as DaNet, DeepLabv3+ and Eanet, and the ablation experiment also verifies the effectiveness of each module of the network in this paper.

Before lung cancer forms a tumor, it often appears in the form of lung nodules. Therefore, a correct diagnosis of lung nodules in time is of great significance to improve the survival rate of patients. This paper proposes a segmentation method of pulmonary nodules in CT images based on MA~~MultiResUnet to assist doctors in diagnosing lung nodules. This method further obtained the feature map with multi-scale spatial information and prominent important channel features by redefining the skip connection structure in the model, and the channel attention module was introduced into the decoder to perform feature calibration, so as to improve the network's segmentation performance of lung nodules. The dataset used the LIDC-IDRI public dataset, and the proposed method was evaluated on the preprocessed dataset. The experimental results show that the Recall, Dice and MIoU performance of MA~~MultiResUnet reach 85.76%, 84.24%, 86.99%, respectively, and the segmentation performance is better than the existing ones.

In order to eliminate the influence of noise on sonar image and improve the recognition quality of sonar image, a method of dual-density dual-tree complex wavelet transform (DDDTCWT) and fuzzy optimization is proposed. The method eliminated the defect of losing practical information due to interval sampling by dual-tree structure, adopted the improved fuzzy optimization algorithm for the decomposed low-frequency components, and performed the dualvariable contraction function processing for the high-frequency components. The enhanced image was obtained through inverse wavelet transform. Experiments show that the proposed algorithm can better retain the details of the image, can make the image level more clear, and has a very good visual effect. The subjective effect and objective index of this algorithm are obviously better than other algorithms.

A hybrid image denoising method based on an adaptive pulsed-couple neural network (PCNN) optimized by bidimensional empirical mode decomposition (BEMD) and the grey wolf optimization (GWO) is proposed. The BEMD decomposed an original image into various bidimensional intrinsic mode functions and a residual, and the decomposed components would be denoised by PCNN optimized with GWO, respectively. The wolf pack algorithm was used to optimize the PCNN parameters. A denoised image was obtained after reconstructing the denoised components. The advantages of this method include: (1) Deter-mining the key parameters of PCNN effectively and improving the convergence speed of the model; (2) Effectively solving the problem of high intensity noise suppression; (3) Preserving the details of the original image completely by isolating the noise points and recovering the original pixels.

At present, normalized body temperature monitoring is implemented in indoor public places. The existing temperature measurement solutions have disadvantages such as slow temperature measurement speed, low temperature measurement accuracy, and small monitoring range. In view of the existing problems, this paper proposes an improved target detection algorithm based on YOLOv5, which is used with binocular cameras to monitor pedestrian body temperature in real time. The algorithm introduced DenseFuse to fuse the input visible light and infrared images at the feature level to obtain feature information of different meanings and enhance the feature structure. The Decoupled Head was used to replace the original coupled detection head to enhance the expression ability of the output and improve the detection accuracy. The experimental results show that compared with the original YOLOv5, the recall rate of the proposed method in this paper is increased by 6.29 percentage points, and the average accuracy rate is increased by 6.37 percentage points, which can meet the needs of efficient and accurate real-time detection in large passenger flow scenarios.

Aimed at the problem of modulation recognition in wireless communication, a lightweight CNN modulation recognition algorithm based on convolutional neural network (CNN) is proposed. The spatial features of the signal were extracted by continuous multiple convolution operations, and features were dimensionally mapped by fully connected layer. The recognition probability was output by using the Softmax layer to achieve the recognition of multiple modulation signal. The trained lightweight CNN model was compressed by X-CUBE-AI and deployed on the STM32F405RGT6 embedded microcontroller, and the RADIOML2016.10a data set was used to test the overall performance of the deployed model. The experimental results show that the designed lightweight CNN model only occupies 1474.6 KiB Flash and 150 KiB RAM, which has fewer parameters than other deep learning models. When the signal-to-noise ratio is 0 dB and above, the recognition accuracy is up to 81.8%. The model can achieve modulation recognition effect comparable to PC platform on ARM Cortex-M embedded platform, which verifies the effectiveness and feasibility of the algorithm.

Under electricity market price uncertainty, power generators need to provide appropriate generation scheduling strategies to maximize their profits. This study proposes a CVaR-based Wasserstein distributionally robust optimization model to address the self-scheduling problem under price uncertainty. Using optimization duality theory, the model is reformulated into a second-order cone programming problem and solved with a commercial solver (Mosek). Furthermore, a region-partitioning-based approximate model is proposed, which utilizes the alternating direction method of multipliers (ADMM) for distributed computation to improve computational performance. Simulation experiments on three test systems are conducted to validate the effectiveness of the proposed model. The simulation results demonstrate that the model effectively balances risk control and profit maximization and is suitable for solving large-scale selfscheduling problems.

Aimed at the problem that the training cycle of the deep reinforcement learning algorithm is too long when it performs full degree of freedom training for manipulator in 3D environment, a fast training method of deep reinforcement learning for manipulator is proposed. By decomposing the grasping task, the training of the lateral steering gear and the longitudinal steering gear of the manipulator was decoupled, and the solution space was compressed by dimensionality reduction, which simplified the training process while ensuring the execution accuracy of the action. The deep deterministic policy gradient (DDPG) algorithm was improved, and the secondary value estimation was performed on the same batch of samples to delay the updating of the strategy network, supplemented by preferential experience replay, which effectively improves the training efficiency of DDPG algorithm. Experimental results show that the proposed method has the characteristics of low training complexity, high speed and low cost, and the success rate of grasping can reach 98%, which is beneficial to the application and promotion of industrial occasions.

Traditional question answering methodsoften have problems of inefficiency and insufficient use of data information. In order to solve the problems above, in the entity recognition part, adversarial transfer learning was used to integrate the boundary information of Chinese word segmentation to improve the accuracy of entity recognition. At the same time, an entity labeling method was proposed based on global pointers instead of CRF to improve model training efficiency. In the predicate matching part, the Siamese network was used to solve the problem of insufficient semantic expression of sentence vectors obtained by using BERT directly. On the data set NLPCC-2016KBQA, an average F1 value of 85.99% was obtained, indicating the feasibility of this method.

Time series similarity detection plays a critical role in scenarios such as financial data analysis and power data mining. To address the quantization loss issue in existing deep hashing networks for time series, we propose an end-to-end Deep Contrastive Time Series Hash (DCTSH) network. By introducing an adaptive binarization network and hash loss, the method eliminates quantization errors during binary hashing, enabling the model to generate time series hash codes with enhanced expressive effectiveness and generalization capability through end-to-end training. For unlabeled time series data, the negative sample selection in the contrastive learning network is improved via clustering to strengthen time series representation learning. Experimental results on multiple time series datasets demonstrate that DCTSH achieves significantly improved detection accuracy compared to previous methods.

More and more cloud services are shifting from monolithic architecture to microservice architecture. Under the microservice architecture, requests will traverse multiple microservices to form a microservice chain. Multiple microservice chains may compete for resources, and the block of a single microservice may be transmitted to any related microservice chain and result in timeout, thus violating the service level objectives (SLO) requested by the user. This paper proposes a dynamic load balancing algorithm for microservice chains that integrates service mesh and deep reinforcement learning, which can satisfy user expectations as much as possible in the face of dynamically changing loads. Experimental results show that the performance of service-mesh based microservice chain split module is improved by more than 10 times compared with the existing methods, and the load balancing algorithm can reduce SLO violations by at least 46% comparedwith the benchmark method.

For the semantic segmentation network, the following problems exist in the fusion of low-level and highlevel feature in the encoder-decoder: (1) feature extraction in space and channel cannot be synchronized, resulting in feature combinations that cannot obtain global context information; (2) feature fusion cannot be fully utilized low-level and high-level feature images, resulting in blurred semantic boundaries. The global atrous spatial pyramid pooling was designed. This structure not only extracted multi-scale information in space and utilized image information in channels, but also enhanced feature reuse in the encoder stage. A feature fusion attention module was designed to connect low-level and high-level features and new features at different stages in the encoder. Experiments show that the algorithm achieves 77.92% mIoU on the Cityscapes dataset.

Aiming at overfitting training data in deep model caused by too few samples, we propose a few-shot learning method that combines knowledge distillation and transfer learning. In order to improve the feature expression ability of shallow network for small sample images, we designed a multi-generation distillation network structure. A modified transfer learning structure was given to enhance the generalization ability of the network by adjusting few parameters. Multiple classifiers were combined to fuse the networks obtained through distillation and transfer. The experiments on three few-shot standard datasets show that the proposed model can effectively improve the classification ability of the model and make the few-shot prediction results more accurate.

The long-tail problem is very common in recommendation system. It leads to recommending few and homogeneous products. We propose a new recommendation algorithm named GES4RL, which combines graph embedding with side information and reinforcement learning to solve long-tail problem. GES4RL is based on Node2Vec and negative feedback reinforcement learning. It constructs a weighted directed graph of product propagation and uses Node2Vec to learn the embedding of products. We used gated recurrent unit (GRU) to learn user's dynamic preferences and designed a negative feedback reinforcement learning model to generate the best recommendation strategy for long-tail products. Experimental results on User Behavior Dataset provided by TianChi show that the algorithm improves the diversity and hit rate of recommendations significantly.

In recent work, Yan et al. designed an auditing protocol with a designated verifier. Note that, Yan et al.'s protocol suffers from great communication redundancy. In order to reduce it, a flexible and improved scheme is proposed, which tries to "aggregate" the tags of multiple data segments and thus greatly reduces the communication overhead. Meanwhile, we give the security proof based on the computational Diffie-Hellman assumption. A concrete example shows that the communication overhead of the improved protocol can be greatly reduced.

The security level of Internet of Things devices is generally low, which has caused many hidden safety hazards. The IoT botnet is one of them. In response to the above problems, an IoT botnet detection model based on RF-RFECV and LightGBM is designed. The model used RF-RFECV to select features according to the importance of features, and used LightGBM to build the model. The experimental results show that this model is improved to a certain extent compared with the traditional detection algorithm in accuracy, precision, recall rate and F1 score, which is useful for future IoT botnet research.

The private key in blockchain wallet is the only voucher in bitcoin transfer transaction. If the private key is held by one person, there will be some problems such as loss and leakage of the private key. To solve this problem, a multi-party blockchain wallet management scheme is designed. The elliptic curve signature algorithm and threshold scheme were adopted, and a decentralized threshold signature scheme for multi-party blockchain was designed. The scheme could prevent the loss of the private key of a single node or the lack of signatures of individual participants, and was suitable for the application scenarios such as wallet management by many people. The correctness and security of the scheme were proved, and the efficiency of blockchain and the time required for transaction under the scheme were evaluated.

Aiming at the defect that most existing protocols can't realize the mutual authentication among server, electronic tag and reader, this paper proposes a protocol which can realize the tripartite authentication based on R-LWE (learning with errors over ring) cryptosystem. The protocol used ideal lattice to encrypt important private information, and the encryption was mixed with different random numbers at the same time, which not only ensured the security of private information, but also made the message fresh and uncertain. The security analysis, formal proof based on GNY logic and performance analysis of the designed protocol show that the protocol has good security requirements, rigorous reasoning and better performance than its comparison protocol.

Aimed at the problems of user information privacy leakage, repeated voting, voting content leakage, and ballot tampering in electronic voting, an anonymous electronic voting scheme based on attribute certificates is proposed. It used zero-knowledge proof protocol to achieve attribute-based certificate authentication and complete user anonymity, used ElGamal encryption technology and the introduction of a vote management server (VMT) to ensure the fairness of votes, and used double Id (CertId, ballot ID) to ensure votes verifiability. Theoretical and experimental analysis shows that the scheme has the characteristics of anonymity, fairness, and uniqueness, and solves the problems of CA semi-credibility and the collusion between CA and voting agency VI, and can be widely used in various election activities.

According to the lack of the transition from qualitative risks to quantitative risks in the current cyber security risk assessment system, this paper proposed an intruder and asset risk assessment method based on multi-dimensional features. From the aspect of the intruder, this paper designed three dimensions to evaluate each intruder, and obtained the top-threat intruders. From the aspect of assets, this paper designed the dynamic scoring of threat dimensions and vulnerability dimensions. Combining the identification and value assignment of assets, we obtained the top-risk assets. The results of this paper can be used to display the current top-risk intruders and risked systems to security analysts, to maximize the existing defense countermeasure. The result is of theory significance and practical meanings.

In order to encourage small-scale power suppliers to effectively share defense data, a measurement infrastructure defense strategy based on sandbox deployment contract theory was proposed. Firstly, an incentive mechanism contract game theory based on information asymmetry was introduced to model the defense data sharing contract between power suppliers. Then, a sandbox deployment contract theoretical model was proposed to improve the defense efficiency of advanced metering infrastructure, ensure the realization of defense data sharing, and reduce the defense cost of traditional power retailers. Simulation results show that the proposed method can effectively realize defense data sharing under the condition of information asymmetry.

The valve hall of the converter station is the core area of the direct current equipment. The fireproof performance of the valve hall blockage plays an important role in supporting the daily operation of the direct current transmission system. This paper analyzed the fire protection capability of the plugging structure of the high-end valve hall in±1 100 kV UHV converter station. This paper determined the fire hazard source according to the material and working mechanism of the blocking structure of the±1 100 kV high-level valve hall. This paper established a blocking physical simulation model of the±1 100 kV valve hall, and simplified it according to the blocking structure components. The t2 fire source model was used to describe the heat release rate of the fire hazard source. The temperature variation trend and the fire spread of the valve side casing when the valve side was on fire were calculated through the visual operation software PyroSim. The simulation results show that the fire spread is limited to the valve side casing area, indicating that the high-end valve hall plugging structure can inhibit the spread of the fire, which provides key basis for designing a reasonable fire extinguishing plan.

The existing processes, techniques, and technologies utilized in the auction system to process the supply chain of commodities are inefficient. They face several challenges in terms of security, dependability, non-modification, traceability, and operational transparency. They also have issues with centralization. Traditional supply chain systems are centralized, vulnerable to single points of failure, and provide ambiguous product information. A decentralized blockchain-based solution was proposed to provide new ideas for the supply chain management process of auction system commodities in a totally secure, transparent, unmodifiable, traceable, and trustworthy manner. This paper combined Ethereum and the interplanetary file system (IPFS) to safely and reliably collect, store, and communicate data pertaining to the auction system's commodity supply chain.