Artificial Intelligence in China

For traffic flow forecasting task exists for the future traffic flow prediction accuracy is low, lack of generalization and the deep learning model and such problems as incomplete, is put forward based on the two-way LSTM traffic flow prediction model of ant colony algorithm, ant colony algorithm for Bi - LSTM network layers, the number of neurons, batch size, number of training to optimize the parameters. In this paper, experiments are carried out on two public data sets of daily traffic flow in Bao‘an District published by California Highway and Shenzhen government open platform. RMSE and MAE are taken as evaluation indexes. The results show that ACO-BI-LSTM model has strong optimization ability and better prediction performance.

This paper analyses the research and application of Natural Language Processing (NLP) technology in the management of construction projects, aiming at the problems such as the expansion of the scale of construction projects and the increasing number of document information, the difficulty and complexity of their management, the inefficiency and complexity of traditional manual collation and induction, the inability of technicians to fully, timely and efficiently consult, modify and improve document information, resulting in the degradation of project performance. Convert unstructured documents in the field of architecture into structured information to solve common problems in architecture.

With the rapid development of industry 4.0 technology, the demand for real-time 3D rendering technology in all walks of life is unprecedented, especially in architectural design and planning, In view of the problems existing in traditional design planning, such as creativity is difficult to express immediately and intuitively, 2D drawings are complex and difficult to understand, workflow is chaotic, and it is difficult for all parties to coordinate and review, This paper analyzes the real-time seamless collaborative creation of SketchUp and twinmotion as a reference to solve the above problems, and uses three cases to analyze the media output from preliminary conceptual design to film and television photo level, in order to build a concise architectural visualization workflow in the spirit of open innovation.

In order to improve the efficiency and quality of software testing, aiming at various factors affecting software reliability, how to find defective modules and optimize them in the early stage of software development has become an urgent problem to be solved, This paper introduces the software defect prediction technology based on life cycle. According to the measurement elements affecting software reliability, relevant internal indicators and design defects, find the defect module, lock it in advance, adopt machine learning technology and reasonably allocate limited resources, which is conducive to evaluate the software design scheme, optimize the design strategy, reduce design changes and improve the software operation process, It plays a role in cost evaluation, resource management, scheme determination and quality prediction in software management. It is hoped to provide some theoretical support and practical reference for the development of software defect prediction.

In order to solve the problem of network attack by malicious code using Domain Name System (DNS), on the basis of analyzing the characteristics of malicious code lines and abnormal operation behaviors, this paper proposes an unsupervised abnormal IP detection method based on DNS log data. Through the construction of DNS fingerprint characteristics, it is used to demonstrate the DNS behavior characteristics of IP to the greatest extent. The detection model is constructed by using isolated forest and local outlier factor algorithm, and the anomaly score of IP is obtained. The experimental results show that the detection method designed in the paper can well detect the attack exceptions and operation exceptions in the network environment. With the help of whitelist, the accuracy of the method can reach more than 90% after selecting the appropriate anomaly score threshold.

Hyperdimensional computing uses hypervectors as basic patterns to construct cognitive codes to represent atomic entities through encodes different types of data into the same data structure based on hyperspace. In this paper, we exploit the reversibility of binary hyperdimensional computing to encode images to hypervectors and decode them back. We introduce turnover rate to properly separate the distance between adjacent values while maintaining the distance between them, so as to avoid the poor effect of segmentation or the direct generation that leads to the distance between adjacent hypervectors being too close to distinguish. We compared the performance of reversibility with the original hyperdimensional computing. The proposed approach has better performance.

Smart water construction is not only an important part of the development process of smart city, but also a key link in the implementation of urban people's livelihood guarantee. Urban water supply dispatching realizes artificial intelligence, which can carry out real-time dynamic control and ensure the stability of the overall water supply pattern. At the same time, it is also the primary platform for the most direct business connection between production technology management and grass-roots units. It also establishes channels for various sections in the field of urban water supply, such as engineering management, measurement management, water quality management, water plant technical measures management and so on. With the continuous innovation and development of urban water supply dispatching, it has experienced the traditional manual experience stage, the primary information stage of production process monitoring and data transmission. At present, we are in the artificial intelligence construction stage including SCADA system, pipe network geographic GIS system and hydraulic model system as comprehensive dispatching auxiliary means.

Aiming at the problems of complex background, serious illumination and small target in optical remote sensing image, a remote sensing image object detection algorithm based on decoupling head and attention mechanism (DHA-SSD) is proposed. The algorithm is based on the combination of resnet-50 and SSD algorithm. Firstly, the decoupling head structure is added to decouple the classification task and positioning task of object detection, so as to alleviate the conflict between the two tasks; Secondly, in the multi-scale detection stage, SimAM attention module is introduced to improve the multi-scale detection performance of the model without adding parameters; Finally, the experimental results on the public RSOD optical remote sensing image data set show that the mAP value is improved by 4.4% compared with the benchmark algorithm.

With the development of smart water conservancy construction, a realistic demand for using computer vision technology to assist in the supervision of floating objects on the water surface has arisen. To address the problem that the current research on water surface floating objects detection mainly focuses on detecting scattered individual floating objects, and an improved YOLOv3 algorithm embedded with the SE attention module is proposed for detecting aggregated water surface floating objects. A self-made dataset containing aggregated floating objects “mixed garbage” and “water pollution” is developed and augmented with four data enhancement methods. The K-means++ algorithm was used to replace the K-means algorithm for clustering the dataset with ground truth box sizes to reduce the negative effects of randomly selecting the initial clustering centers. The localization loss in the loss function of the YOLOv3 model is improved, and GIoU Loss is introduced to improve the localization accuracy. The experimental results show that S-YOLOv3 outperforms other models in the field of water surface object detection on the self-made dataset compared with YOLOv3 and other commonly used models in the field of water surface object detection, and the mAP reaches 83.1%.

In recent years, natural language generation of SQL sentences (Text2SQL) has received a lot of attention as an important research direction natural language processing (NLP). Text2SQL makes it easier for users to query complex databases without learning SQL sentences and the underlying database schema. The current mainstream Text2SQL method is the grammar-based IRNet, which attempts to present an efficient method with explanations from the perspective of constructing reasonable grammars and provides a good solution for solving complex nested queries. Still, it makes simple use of external database ontology knowledge, resulting in natural language problems in which words do not correspond well to tables and columns in the database. To address this problem, a new method that considers the entity relationships between natural language problems and data in the database - SLESQL is proposed, which extends some of the functionality of IRNet by using schema linking enhanced Experiments show that SLESQL achieves 6.8% improvement in accuracy over IRNet on the publicly available dataset Spider.

Emotion-Cause joint extraction is to extracting both the emotion and its corresponding cause from the given text, which has a wide range of application scenarios. Previous work only considered emotion extraction, cause extraction and emotion-cause relation classification tasks. This paper proposes a multi-task learning model based on emotion classification to perform emotion-cause joint extraction in a unified model, which obtains semantic features of different granularity based on Bert and attention mechanism. We also introduce focus loss function to deal with the sample imbalance problem. Experimental results show that the emotion classification subtask can effectively extract the emotional features of documents. Compared with the existing models, the experimental results show that our model outperforms the state-of-the-art model on emotion-cause joint extraction.

The construction environment is complex and dangerous, and it is difficult to achieve all-round, whole-process and real-time helmet detection. In order to ensure the safety of personnel, this paper proposes a helmet detection algorithm based on improved YOLO v5s. First, replace the backbone feature extraction network of YOLOv5s with the end-side neural network architecture GhostNet, which greatly reduces the amount of network parameters. Introduce the lightweight module attention mechanism ECA-Net in the C3 module to improve the feature extraction ability, and finally use the CIOU as the loss function to improve the positioning accuracy. The average accuracy (mAP) of the improved model on the SHWD dataset reaches 93.86%, and the processing speed (FPS) reaches 49. Compared with the original YOLOv5s, the amount of parameters is reduced by 13.33% without reducing the mAP, and the size of the model is reduced. 26.6%, processing speed increased by 22.5%. The experimental results show that it can effectively reduce the amount and size of model parameters and meet the real-time detection requirements of embedded devices.

Automatic essay scoring techniques can automatically evaluate and score essays, and they have become one of the hot issues in the application of natural language processing techniques in education. Current automatic essay scoring methods often use large pre-trained models to obtain semantic features, which do not perform well in the field of automatic essay scoring because the training corpus does not match the content domain of the essay, and the extraction of features for long essays is not effective. We propose a label embedding-based method for scoring secondary school essays, using a modified bidirectional long- and short-term memory network and a BERT model to extract domain features and abstract features of essays, while using a gating mechanism to adjust the influence of both types of features on essay scoring, and finally automatic scoring of essays through feature fusion. The experimental results show that the proposed model performs significantly better on the essay auto-scoring dataset of the Kaggle ASAP competition, with an average QWK value of 81.22%, which verifies the effectiveness of the proposed algorithm in the essay auto-scoring task.

Wheat ears detection and counting play a crucial role in wheat yield prediction and breeding. In this paper, a deep neural network wheat ears detection method Wheat-YOLOv5 based on improved YOLOv5 is proposed for the problem of low accuracy of traditional wheat ears detection methods. Fusion of the ECANet attention module with the Backbone part of the YOLOv5s network to improve network feature extraction. Using SPConv convolution to replace the original ordinary convolution in the neck layer to improve the model’s ability to cope with complex scenes of wheat ears. Using $$\alpha $$ α EIoU Loss instead of GIoU Loss as the target bounding box regression loss function to improve the accuracy of wheat ears localization. The detection average accuracy AP value of the improved algorithm reaches 94.30% and F1 value reaches 91.50%, which has certain recognition accuracy and robustness and can effectively improve the detection of wheat ears in actual agricultural scenes.

Aiming at the characteristics of multi-scale, diversity and complex background of aircraft targets, in order to improve the average detection accuracy of YOLOv5s algorithm, an improved aircraft target detection algorithm based on YOLOv5s model is proposed, Firstly, replace conv module in the backbone network of YOLOv5s with RepVGGBlock to reduce the number of parameters; Secondly, a small target detection head is added to enhance the recognition ability of small targets; Finally, GAM_Attention mechanism is introduced in front of each detection head to improve the detection accuracy. The research shows that the RG-YOLOv5s (Repvggblock GAM_Attention you only look once) algorithm proposed in this paper improves the average accuracy by about 1% and reaches 97.3% when IOU = 0.5, which is more suitable for the detection of aircraft remote sensing targets.

With the rapid development of the metaverse, this is the inevitable trend content form from 2D to 3D, from real to virtual to virtual-real combination. The demand for virtual scenes will surge in the coming period. How to quickly generate a large number of high-quality scene graphs has sparked extensive academic attention. GAN being the most popular generation algorithm, and we discuss the prevalent architectures to solve the problems of training collapse and uncontrollable images during training. Besides, this paper focuses on the development of GAN and its variants in the field of image synthesis. A comprehensive review of recent advances in virtual image generation in the last two years is presented, with a summary of the classification of algorithms and an introduction to techniques related to multimodal generation. Finally, some suggestions for future research directions are proposed.

The development of modern science and technology has brought us a far-reaching and significant impact. Artificial intelligence is an important product of this era. With the continuous development of society, the status of social vulnerable groups has been paid more and more attention. As a country with a large population, China has the highest number of disabled people in the world. The existing products and services for the disabled can not meet the needs of users. The application of artificial intelligence technology in interior design and product design practice is still in the exploratory stage, and there is a large space for development. Through the user research of observation method and experience prototype method, this paper deeply excavates the communication needs and expectations of Chinese disabled people. Through user research and design, a set of smart home technology service system is proposed, which provides a reasonable and feasible solution for the daily communication of the disabled.

With the rapid development of artificial intelligence, the application of Computer vision and Machine learning in the field of photogrammetry and remote sensing continues to enrich. Cognitive reasoning based on spatiotemporal big data has gradually deepened, which greatly promotes the development of remote sensing and surveying and mapping geographic information technology. The natural resource survey and monitoring technology system presents the characteristics of intelligence, spatialization, ubiquity and multi-source, which promotes the development of natural resource survey and monitoring business.

Natural resources are the material basis for human survival and national economic and social development. Natural resources survey is related to major national decision-making and deployment, fully supports and serves the economic and social development and ecological civilization construction in the new era, and systematically grasps the quantity, quality, spatial distribution, development and utilization, ecological status, and dynamic changes of national natural resources and landmark substrates. Computer vision technology uses computers to process, analyze and understand images to identify targets and objects in various patterns. The natural resources are photographed and imaged by industrial unmanned cameras or remote sensing satellites, and the characteristics are compared according to image processing, to realize the investigation and analysis of natural resources, improve the efficiency of natural resource management investigation, and provide a reference for planning decision-making and implementation.

Generative Adversarial Network (GAN) has become a research focus in the field of deep learning, and its research output has grown exponentially. This brand-new technology provides new ideas and methods for object detection, and has achieved remarkable success. Firstly, this paper introduces the basic GAN model and its derivative models in the field of object detection. Then analyzes the application status of GAN from object detection fields, such as industrial defect detection, medical image detection, remote sensing image detection, and face detection. Finally, summarize and prospect the technology development of generative adversarial networks.

Convolutional neural networks have become one of the important research directions in the field of computer vision based on deep learning, and they are widely used in object detection with excellent performance. Based on the research results of many scholars, the paper reviews the application research of convolutional neural networks in object detection, introduces its application research from three aspects: object detection based on candidate region, object detection based on regression and video object detection, and finally summarizes the development of convolutional neural networks in the field of object detection.

This paper proposed an optimal deployment algorithm for large-scale wireless sensor networks based on graph topology. The algorithm mines the intrinsic property of sensor measurement data as node quality through auto encoder and defines topological potential to describe the relationship between sensor nodes as the distribution of the topological potential field, thereby dividing the large-scale graph into several subgraphs. Next, singular value-QR decomposition is used to find the crucial nodes in each sub-network, and achieve the optimal deployment of large-scale sensor networks. We experiment with this algorithm on the CIMIS dataset. The results show that the mean square error of this algorithm is only 0.02116.

Structured Query Language (SQL) is a query language widely used in databases, Text2SQL automatically parses natural language into SQL, which has great potential to facilitate non-expert users to query and mine structured data using natural language. Current research focuses on improving the matching accuracy of SQL clause tasks, but ignores the correctness of SQL syntax generation, and SQL generation involving multi-table joins still suffers from a large number of errors. Therefore, a neural network-based Text2SQL approach is proposed. To implement a practical Text2SQL workflow, the model associates natural language queries with an inverse normalized database schema, called INSL (Inverse Normalized Schema Link Generation Network). Through theoretical analysis and experimental validation on the public dataset Spider, INSL can effectively improve the quality of Text2SQL tasks.

In recent years, China’s auto manufacturing industry has been developing rapidly and Chinese car companies are facing both internal and external pressures. As a result, managers of automobile enterprises are now paying more attention to the financial situation of their enterprises, hoping to avoid the risk of bankruptcy by predicting financial distress in advance. Most of the existing research findings are based on financial index data or annual financial reports to build quantitative and qualitative models to predict the risk of financial distress of automobile enterprises. This study innovatively proposes to combine financial indicators data, financial text data and non-financial text data to predict the risk of financial distress of automobile enterprises, and constructs a set of prediction models based on deep learning algorithms, which is a useful attempt of cross-disciplinary research in this field. In terms of experimental analysis, we found that the introduction of financial text data and non-financial text data into the model significantly improves the prediction performance compared with the traditional prediction model based on financial indicator data, which indicates that the combination of online user reviews and financial annual reports can be more helpful for predicting the financial distress risk of car companies.

Rain and snow contain particles generated by different physical and chemical processes. Weather radars send directional pulses of microwave radiation, on the order of a microsecond long. Between each pulse, the radar station serves as a receiver as it listens for return signals from particles in the air. This is the measurement principle of microwave weather (meteorological) radar. Limited by the information processing capability, the weather radar data published by the meteorological website often have a large time interval, such as 6 min. Video frame interpolation technology has made great progress in recent years with the development of deep learning technology. The frame interpolation of weather radar charts will bring users more accurate weather descriptions and more intuitive decision-making references. This paper propose a novel video frame interpolation algorithm for weather radar data, which is able to generate the intermediate frames between the frames at the sampling time.

Currently, many deep learning based methods for music generation have been proposed. However, these methods use time-domain audio data to train their networks; consequently, these methods generate low-quality music. This paper proposes using frequency-domain data that are transformed from time-domain data to train a hybrid neural network that combines a recurrent neural network with a generative adversarial network to generate music in an end-to-end way. The automatic music generation method proposed in this paper explores a new representation of music data, and the results prove the effectiveness of our method.

Models in previous studies about inclusive finance often include economic data while excludes public online statements. In this paper Random Forest Regression (RFR) model is trained on the annual influencing factors and annual financial inclusion index to predict quarterly financial inclusion index by the quarterly influencing factors to expand the size of data. Then, BOW model tf-idf algorithm is used to convert COVID-19 – loan related online statements into word vectors. Lastly, these influencing factors of different lag periods are passed into the RFR model to compare their performance. Result of models shows that there is impact the epidemic has on the development of inclusive finance, and the lag period of the impact opinion texts on financial inclusion is 2 quarters.

Visual pose estimation is essential for many space applications, such as autonomous robots for planetary exploration, asteroid exploration, and spacecraft pose estimation. However, the high reliability requirement of space applications cannot be satisfied by general visual pose estimation algorithms based on heuristic local optimization. This paper briefly introduce certifiable optimal algorithms for the visual pose estimation framework to the aerospace research community, including both visual relative pose estimation problem and pose-graph optimization problem. The original optimization problem is reformulated as a quadratically constrained quadratic program (QCQP) problem, whose Lagrangian dual function is a semidefinite programming (SDP) problem which is convex. Simulation results demonstrate the feasibility of certifiable optimal visual pose estimation algorithms based on the zero duality gap for typical application scenarios.

This research proposes a CNN-based model for automotive manufacturing companies to predict quarterly revenue that not only combines the sentiment variables of online reviews with financial indicators, but also covers four-quarter indicators of the firm itself, the industry average, and the average of the companies before and after their ranking. According to the experimental findings, our suggested model outperforms the conventional machine learning model Random Forest in terms of prediction accuracy by 75.2%. Business revenue predictions that use a sample of data from three dimensions, including the company and the industry average, perform 48.7% better than those that use only the company’s own data. The model with comment variables outperforms the model without comment sentiment variables in terms of prediction performance.

In the conventional iterative blind source separation procedure, the nonlinear function is selected as a fixed mathematical expression that cannot change with different signals. In this paper, we propose an enhanced independent component analysis (ICA) network for sub-Gaussian sources. The basic principle of the network is to replace the nonlinear function with an explainable neural network, and then to incorporate deep learning into the iterative computation for optimization. To give the explainable neural network a certain function expression ability, we design its structure and pre-train it. The gradient descent approach then be used to update the weights of the enhanced ICA network. To illustrate the effectiveness of the algorithm, we merge four different signals and perform separation experiments. Experimental results indicate that the enhanced ICA network can be consistent with the performance of the iterative algorithm, and in such conditions it can obtain even more precise waveforms.

Correct and complete aviation data is a prerequisite for SWIM cross-domain information sharing, which will affect the correct operation of the application program of the air traffic control system. The current syntax-based data verification method can verify the good structure of AIXM exchange documents. And rule constraints are usually in raw textual form and cannot be automatically enforced in computerized systems. This paper proposes an automated data semantic verification method for business rules based on SBVR, constructs the semantic rules meta-model of AIXM data, forms a set of AIXM business rules writing methods, and realizes the coding conversion path for implementing this method into system development, which supports automatic verification of more complex business logic and rule constraints.

Data-driven methods are broadly used in analyzing big data in many fields. ADS-B trajectory data provided the possibility of anomaly detection in terminal airspace. Approaching and landing accidents are usually originated from unstable approaches, which are the consequence of go-around failure. Go-around is an aborted landing process initiated in the event of an unsafe landing. In this paper, DBSCAN and HDBSCAN clustering algorithms were employed to detect spatial and energy anomalies using ADS-B trajectory data. Air traffic flow patterns are identified, thus energy safety boundaries are established. Finally, atypical scores are quantified based on the total energy, which serves as the foundation for studying, comprehending, and addressing the factors that cause unstable approach events and go-arounds. This method detects the unstable approach from the spatial and energy perspectives and quantifies the degree of anomaly. It is essential for detecting and predicting anomalies in terminal airspace, and also explores the application of data-driven methods to aviation data.

Pedestrian detection is a technology that uses computer vision to determine whether there are pedestrians passing by in the video sequence or pictures, and realizes the positioning of pedestrians. It is an important task in manless driving, automobile intelligence and intelligent monitoring. Aiming at the problems of low efficiency of pedestrian detection and slow running speed on small devices, a YOLOV6-SE (Yolo Look Only Once-SE) model is constructed to detect pedestrian targets。the mAP detected by pedestrians reaches 85.1%, which is 7.3% points higher than that of the original YOLOv6 without adding attention mechanism. The three-layer channel attention module is added to backbone(RepVGG) adopted by YOLOv6, which enables the backbone network to extract more feature information, thus improving the accuracy of the network. After experimental comparison, Good detection performance is achieved.

Computer Vision is one of the popular research in the field of Deep Learning. As a hotspot in the field of computer vision research in recent years, pedestrian detection has a wide application prospect and great economic value in urban transportation, social public safety, national defense security construction, and so on. In this paper, the Yolov4-Tiny algorithm is used as the basic algorithm to implement pedestrian detection in crowded subway stations, and the network structure, detection principle, and training process of the algorithm are studied in detail. The map value achieved by the model after training is 84.42%, and pedestrian images are selected and then tested with good results. In this paper, the trained model is tested on the embedded motherboard Jetson Nano, and the processing speed fps value is about 7, which can meet the requirements of real-time pedestrian detection.

Coronary heart disease has become a disease with the highest mortality rate in the world. The main treatment for coronary artery disease is coronary stenting. The detection and reconstruction of stents after coronary stent implantation is currently a difficult problem. In this paper, a reconstruction method for coronary stent detection based on IVOCT images was proposed. This method utilized the continuity of IVOCT image frames and the columnar shadow of the stent to realize two-dimensional and three-dimensional reconstruction of the stent. Median filtering, Otsu method, Canny method, pixel fetching and coordinate conversion were used. It was verified that the reconstruction of stent images could be achieved accurately when the frame spacing of IVOCT images satisfies the sampling theory. The stent recognition reconstruction method could be used in clinical applications to assist physicians in post-stenting status analysis.

In this paper, we address a concept drift detection and update algorithm based on the online restricted Boltzmann machine (O-RBM). We introduce an attention mechanism into RBM, and the parameters of each classifier in the concept drift detection model are updated according to the important information mined by the attention mechanism. The updated model complies with the various data better and judges the types and states of new data online. In the experiments, we compare the performance of the proposed algorithm with the algorithm proposed in our previous work, and the results show that O-RBM plays even better in concept drift detection.

In recent years, artificial intelligence technology has made breakthrough progress. It is widely used in person re-identification(Re-ID), but its accuracy still needs to be improved. This paper proposes a person re-identification(Re-ID) model based on 3D data augmentation and designs and implements edge computing deployment. The 3D human body model is put into the 3D scene and take automatic virtual photography to obtain the simulation image data. The neural network model is trained through the expanded data set of 3D simulation, and its improvement is significant. Based on this, an onboard scheme of artificial intelligence edge computing based on the Rock chips 3588 was designed to meet the practical needs of person re-identification(Re-ID) applications. This scheme is widely compatible with standard IP cameras based on RTSP/RTMP streams, supports person re-identification of visible and infrared video streams, can control the PTZ through the network interface, and can be updated and iterated remotely through the network.

With the rapid development of 5G, telecom operators have begun to deploy and promote 5G services. How to accurately recognize 5G potential customers and increase conversion rate of 5G customer has become a very concerned issue for telecom operators. In this paper, a 5G potential customer recognition model based on telecom big data is constructed. Firstly, a influencing factor system is constructed to fully reflect the relevant characteristics of 5G customers. Secondly, feature selection is performed by summing the importance evaluated by multiple tree ensemble models, and it is found that relatively important features include data ARPU, brand type, DOU and its changes, ARPU and its changes, video APP usage and so on. Thirdly, the SMOTE-ENN combined algorithm is applied to improve the performance of the basic model. Then, various machine learning models are used to identify 5G potential customers, and the comparison results show that XGBoost, LightGBM and DNN have better performance. Finally, based on the three best-performing classifiers, a multi-layer heterogeneous integrated model is constructed, which combines the cascade and parallel structures. Experiments show that the MHI model can achieve better performance than the separate classifiers.

With the popularization of artificial intelligence, 5G and other technologies, a number of emerging applications require both efficient communication and computing service, which poses enormous challenges to the computing ability and battery capacity of terminal equipment. Moreover, ground-based 5G system can not provide seamless service especially for hotspot and remote area. To tackle the above challenges, we minimize the weighting of delay and energy consumption by optimizing the task offloading decision and computing resource allocation, which, however, is a mixed integer nonlinear programming (MINLP) issue due to the strong coupling between optimization variables. Therefore, we decompose it into two subproblems and design a deep reinforcement learning-Based approach to address the first problem with offloading decision-making. For the second computing resource allocation subproblem,. a Greedy-based solution is proposed. The simulation results indicate that, in comparison to other benchmark approaches, the proposed method can achieve superior performance.

In this paper, we investigate a deep Q-network (DQN)-based method for applying a dynamic spectrum access model to device-to-device (D2D) communications underlying cellular networks. Dynamic spectrum access (DSA) devices have two critical concerns, namely avoiding interference to primary users (PUs) and interference coordination with other secondary users (SUs). We consider that the issues faced by DSA users are also applicable to the D2D communication underlying cellular network. Therefore, we propose a distributed dynamic spectrum access scheme based on deep reinforcement learning (DRL). It enables each D2D user to learn a reliable spectrum access policy through imperfect spectrum sensing without knowledge of system prior information, avoiding collisions with cellular users and other D2D users and maximizing system throughput. Finally, the simulation results demonstrate the effectiveness of our proposed dynamic spectrum access scheme.

In recent years, many power allocation algorithms to maximize energy efficiency (EE) have emerged in wireless communication systems (WCS), but these traditional power allocation algorithms have high computational complexity. The advanced deep learning technique proposed in this paper is shown to solve the transmission power control problem in wireless networks to optimize EE. From a machine learning perspective, the conventional power allocation algorithms can be viewed as a nonlinear mapping between channel gains among users and the optimal power allocation scheme, and deep neural network (DNN) can be trained to learn this nonlinear mapping. Based on this, a DNN-based power allocation method is proposed, and the specific structure of the DNN and the system model of the DNN method are introduced to maximize the EE among users in WCS. The results show that the performance of the proposed method using DNN is essentially the same as that achieved by the conventional algorithm, but the computational time is greatly reduced.

A television jumper is a large entertainment project in which visitors ride a jumper to watch mages that correspond in real time, and is a high speed, heavy duty lift. This paper uses virtual prototype technology to analyse and optimise the safety clamp of a television jumper. A 3D solid model of the safety caliper is carried out using the software Pro/E. Then, appropriate constraints and loads are added to create the virtual prototype of the safety clamp. The virtual prototype was simulated to obtain the braking performance curve of the safety clamp, and the braking performance of the safety caliper was analysed. This paper gives the core algorithm GISTIFF. The parametric optimisation method provided by ADAMS is used to optimise the preload and contact force material of the safety clamp, providing a reference method for the commissioning and installation of the safety clamp as well as improving the design safety.

The analysis and forecasting of non-stationary financial data with artificial intelligence techniques are useful for comprehending the future economic climate and have become a hot research topic. In this study, we propose a novel approach combining variational mode decomposition (VMD) and temporal convolutional network (TCN) for single-dimensional financial time series forecasting. First, VMD is used to reduce the influence of random noise on the prediction model. After that, we utilize TCN to learn the temporal dependence of subsequences. Additionally, we also compare the effectiveness of VMD and EMD for feature extraction of financial time series. The impact of the modal quantity K’s value is also evaluated when VMD is used to analyze financial data. We validate the proposed method in a broad experimental analysis over 3 publicly available datasets. Experimental results show that our method achieves significant improvements in the prediction accuracy of financial data.

In view of the development of low-orbit satellite constellation and the increasing demand for computing ability and privacy protection of space-based applications, a satellite-based multi-task federated learning system is designed. Specifically, with consideration of both the fluctuations of satellite-to-ground communication links and local training impacts on global models, a user selection strategy is established. On this basis, the functional relationship between task convergence speed and wireless communication factors is analyzed, and the KM optimal matching algorithm is adopted to dynamically allocate communication resources. Simulation results show that our proposed scheme outperforms baseline methods in both total weighted error rate and accuracy, and achieves the fastest convergence rate at the same time.

In practical industrial scenarios, efficient anomaly detection is important for the development of industrial system safety and maintenance. In this paper, an industrial time-series signal anomaly detection algorithm based on Gaussian confidence interval and long short-term memory auto-encoder (G-LSTM-AE) is proposed to improve the accuracy and slow detection efficiency of traditional industrial time-series signal anomaly detection. The proposed algorithm is consisted of four steps: time-series data preprocessing, G-LSTM-AE model constructing, Gaussian confidence interval solving, and time-series signal real-time detection. To verify the effectiveness of our proposed algorithm, we conduct the experiments on automatic guided vehicle (AGV) dataset and Skoltech Anomaly Benchmark (SKAB) dataset, respectively. Compared with the existing unsupervised time-series signal anomaly detection algorithm, the accuracy and recall rate of the proposed G-LSTM-AE based algorithm achieves 95% (3–5% improvement). Meanwhile, the Gaussian confidence interval anomaly analysis method improves the reliability and practicality of anomaly detection in industrial scenarios.

In order to effectively reveal the various scientific research entities and semantic relationships among them in the field of hydrogen energy technology, the top-down construction method was used to explore the construction process of the scientific research knowledge graph in the field of hydrogen energy technology. Ontology was used to construct the schema layer of the knowledge graph, after knowledge extraction and knowledge fusion in the data layer, the knowledge was stored in the Neo4j graph database. The knowledge graph constructed included 345300 entities of 8 types and 2167484 entity relationships of 12 types. Through the construction of the knowledge graph, the complex knowledge system visual analysis of various scientific research entities and their relationships can be effectively realized, it can provide support for researchers to grasp the whole research situation in the field of hydrogen energy technology.

This paper analyzes the problems in the development of foreign language education under the background of language intelligence, and proposes solutions to these problems. This paper holds that under the background of the great development of language intelligence, foreign language education should construct the smart foreign language education system by improving the construction of the teaching staff, constructing the teaching research and teaching management platform, and realizing the benign development of foreign language education.

In recent years, artificial intelligence (AI), which advocates machine learning, has attracted the attention of many educational scholars with its unique deep learning technology. With the development of artificial intelligence technology, the research of linguistic intelligence has made remarkable achievements. These advances not only promote the development of intelligent foreign language education, but also pose challenges to traditional foreign language education in China. This paper will make a preliminary analysis of the current situation and development prospect of AI-empowered foreign language education in China.

We briefly review complex-valued neural networks (CVNNs) and compare them to real-valued neural networks (RVNNs). CVNNs allow for a richer representation of many wavelike based physics and engineering fields by retaining the data structure and correlation between real and imaginary parts of the signal. For example, most RF modulations use in-phase and quadrature components in which analysis benefits from the use of CVNNs. We then present state of the art in wireless radio applications utilizing CVNNs and/or complex-valued data. Wireless applications reviewed include modulation recognition, signal identification, channel sensing information and specific emitter identification. Finally, we present motivation for future exploration of CVNNs.