Subjects: Information Science and Systems Science >> Control science and technology submitted time 2024-01-13
Abstract： As a new field with rapid development in the past ten years, deep learning has attracted more and more researchers' attention. It has obvious advantages compared with shallow model in feature extraction and modeling. Deep learning is good at mining increasingly abstract feature representations from raw input data, and these representations have good generalization ability. It overcomes some of the problems in AI that were considered difficult to solve in the past. With the significant increase in the number of training data sets and the surge in chip processing power, it has achieved remarkable results in the fields of target detection and computer vision, natural language processing, speech recognition and semantic analysis, so it also promotes the development of artificial intelligence. Deep learning is a hierarchical machine learning method that includes multilevel nonlinear transformations. Firstly, this paper discusses the basic knowledge of deep learning, analyzes the superiority of the algorithm, and introduces the mainstream learning algorithm and its application status. Finally, the existing problems and development direction are summarized.
Peer Review Status:
Awaiting Review
Subjects: Computer Science >> Computer Application Technology submitted time 2024-01-09
Abstract： Object detection under low illumination conditions is an important task in image processing. The current research pay attention to reduce image noise by image enhancement, improve network structure and data sets to adapt to object detection under low illumination conditions. However, few people have studied the specific influence of low illumination conditions on object detection. Therefore, in this paper, we generate data sets that simulate low illumination conditions through algorithms. Then, we conduct object detection under different noise conditions and collect results, research the impact on object detection.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Other Disciplines of Computer Science submitted time 2024-01-08
Abstract： With the rapid development of sensor and network technology, a large amount of historical time series data appears, so it is more and more important to predict time series efficiently and accurately. In recent years, the methods of applying deep learning ideas and techniques to time series prediction tasks have developed rapidly and achieved many results. This paper analyzes the research status of time series forecasting methods at home and abroad, discusses the relevant theories involved in time series forecasting, summarizes the traditional methods used in this task, the methods based on machine learning and the methods based on deep learning, and focuses on the comparison and analysis of the advantages and disadvantages of each method based on deep learning. Then, the prediction methods of time series based on deep learning are forecasted.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Computer Application Technology submitted time 2024-01-07
Abstract： In this paper, the classification of feature-level data fusion algorithms is summarized, and the distribution is summarized from the fusion algorithm based on probability and statistics, the fusion algorithm based on logical reasoning, the fusion algorithm based on feature extraction, the fusion algorithm based on search and the fusion algorithm based on neural network, and the future research direction of data fusion is summarized.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Natural Language Understanding and Machine Translation submitted time 2024-01-07
Abstract： Multilayer perceptron (MLP) is a feedforward neural network that overcomes the limitations of linear models and opens the door to deep learning by adding one or more hidden layers to the network. In this paper, multilayer perceptrons are used to classfy image, which is explored on the Fashion MNIST dataset, and is attempted to be migrated to the MNIST dataset. In Fashion MNIST, we selected different optimization methods and compared them after feature preprocessing, optimized and improved the multi-layer perceptron by adding regularization methods such as dropout and weight decay.
Experiments show that appropriate feature processing can improve the numerical stability of the model. The momentum method significantly improves the effect of the model, weight decay and other regularization methods help to improve the generalization effect of the model.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2024-01-07
Abstract： Handwritten digit recognition based on deep convolutional networks has been increasingly studied and applied due to the highly nonlinear de In common convolutional neural networks, they are usually composed of input layers, convolutional layers, pooling layers, activation layers, and fully connected layers connected in a certain order. The input layer of the convolutional neural network implements the input of the entire neural network. In this design, the training and inference data is a single channel grayscale image of 30 * 30 pixels
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Other Disciplines of Computer Science submitted time 2024-01-07
Abstract： Image classification and recognition are of great significance in modern society. There have been many excellent convolutional neural network works to optimize the accuracy of image classification, one of the outstanding representatives is ResNet 1 , which greatly increases the depth of the neural network, thereby greatly improving the performance of the neural network. At the same time, there are some pluggable performance optimization sub-modules that can help optimize all networks, one of the outstanding representatives is SeNet 3 . However, they do not always perform well when faced with complex scenarios in the real world. The main work of this article is to study how to effectively improve the recognition performance of convolutional neural networks (ResNet) in some special scenes (small pictures, high-noise pictures), and try to analyze the underlying mechanisms of some neural networks.
Peer Review Status:Awaiting Review
Subjects: Information Science and Systems Science >> Control science and technology submitted time 2024-01-07
Abstract： Traditional motor fault diagnosis technology is usually based on a single type of state parameters, such as vibration parameters or electrical parameters. However, the monitoring range of a single type of motor state parameters is very limited in many cases, which is difficult to meet the needs of comprehensive fault diagnosis of motors. The purpose of this paper is to propose a comprehensive motor fault diagnosis method by fusing vibration data and current data, so as to improve the reliability and accuracy of diagnosis. On the basis of data fusion, it is considered that in the actual industrial and production environment, the cost of obtaining large-scale labeled samples is often high or even not feasible. Therefore, the neural network is further studied and improved, and a small sample fault diagnosis network based on RNN and attention mechanism is proposed.
In this paper, the motor fault feature extraction method is used to study the vibration and current signal characteristics of the motor under different faults. The fault feature extraction methods adopted include Fast Fourier Transform (FFT) and Hilbert-Huang Transform.
According to the actual data fusion requirements in this paper, the overall implementation scheme of data fusion is designed. The fault features are extracted by using FFT, Hilbert-Huang Transform (HHT) and Convolutional Neural Network (CNN) and Multilayer Perceptron (MLP) in turn, and the vibration and current parameters of the motor are fused to carry out comprehensive fault identification and fault diagnosis. The results show that the motor fault diagnosis technology using data fusion method can improve the accuracy of diagnosis results and reduce the uncertainty caused by a single parameter, thus improving the accuracy of motor fault diagnosis. The designed small sample fault diagnosis network is used to identify the health status of equipment under small samples, in which the attention mechanism captures the spatial and channel relationship of signals, and a single experimental sample is used to verify that the network used in this paper has the advantages of diagnostic efficiency and accuracy under different small sample working conditions.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Natural Language Understanding and Machine Translation submitted time 2024-01-07
Abstract： The goals of this Chinese named entity recognition project mainly include the following two aspects. Firstly, it is to achieve high-precision Chinese named entity recognition. Through deep learning of Chinese text, the accuracy of Chinese entity recognition is improved, and the phenomenon of misidentification and missed recognition is reduced. Secondly, it is necessary to establish a standardized process and form a standardized Chinese named entity recognition process, including data preprocessing, model training, entity recognition, etc., to provide a foundation for subsequent research. The code has been submitted to GitHub at https://github.com/Blue88888/DL_CNER .
Peer Review Status:Awaiting Review
Subjects: Electronics and Communication Technology >> Communication submitted time 2024-01-07
Abstract： Configuration verification of networks, especially virtual private networks, is a complex task that needs to be done before every update of a production environment so that network providers can ensure network availability for their customers. This paper discusses a graph-based neural network (GNN) approach for detecting and locating configuration errors in IP virtual private networks (VPNS). The study focuses on two GNN models, one that focuses on routing misconfigurations between customer and provider edge routers, and the other on VPN routing misconfigurations between different provider edge routers. The goal is to provide a tool that simplifies the process of verifying an end-to-end VPN configuration.
In the study, both models were trained using a balanced dataset containing examples of tag configurations extracted from an IMSNetwork-based VPN deployment. The results show that both models show high accuracy when dealing with VPNS of different sizes (from 3 to 40 sites) and two types of architectures (full mesh and hub radiant).
The advantage of this method is that the graph neural network can capture the complex relationship between network topology and configuration, so that configuration errors can be detected more effectively. By using this technology, network providers can validate network configurations before each update to ensure network availability for their customers.
Peer Review Status:Awaiting Review
Subjects: Electronics and Communication Technology >> Information Processing submitted time 2024-01-07
Abstract： In recent years, with the rapid development of deep learning technology, image recognition based on convolutional neural networks (CNN) has achieved remarkable achievements in various fields. In the field of botany, flower species identification is an important research direction and is of great significance in ecology, agriculture, and environmental monitoring. This research aims to explore and optimize the application of ResNet (deep residual network) in flower type recognition tasks. First, the article conducts an in-depth analysis of the structure of the ResNet network, understands its mechanism for introducing residual learning, and how to effectively deal with the vanishing and exploding gradient problems in deep network training. Through preliminary experiments on a large-scale flower image data set, the excellent performance of ResNet in handling complex multi-category flower image recognition tasks was verified. In the data preprocessing stage, the article uses data enhancement techniques, including cropping and flipping, to expand the training data set and improve the generalization ability of the model. At the same time, the flower images are standardized to adapt to the requirements of the ResNet network for input data. Experimental results show that compared with the traditional CNN model, the flower type recognition model using ResNet has significantly improved accuracy and convergence speed. In addition, through in-depth analysis of the model's performance on different flower categories, it was found that the ResNet network performed better when processing flower images with hierarchical structures and complex shapes. The model proposed in this article not only achieves excellent performance overall, but also has high accuracy in identifying specific flower categories. In further research, we consider further improving the generalization ability of the model through transfer learning, especially when facing small sample flower data sets. At the same time, the real-time performance of the model will be explored to adapt to the need for rapid and accurate identification of flower types in real scenes.
This study provides a useful reference and reference for the application of deep learning in the field of botany by conducting a comprehensive and in-depth analysis of the advantages and applications of ResNet network in flower species recognition tasks. The research results not only have certain theoretical value for the improvement of flower identification technology, but also have extensive potential for promotion in practical applications.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Computer Software submitted time 2024-01-07
Abstract： This thesis proposes a method for industrial process control loop fault diagnosis based on graph neural networks. By monitoring the output signals of loop sensors, the graph neural network can capture abnormal behaviors in the loop and automatically diagnose the type of loop faults. Experimental results demonstrate that the proposed method can efficiently detect loop faults and achieve high accuracy in both single and multiple fault scenarios. This method provides a reliable fault diagnosis solution for industrial process control, which has important practical significance and application value in actual industrial applications.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Natural Language Understanding and Machine Translation submitted time 2024-01-07
Abstract： In response to the current problems of inadequate and incomplete semantic feature extraction in Chinese named entity recognition research, Transformers (BERT) has shown striking improvements in a variety of related NLP tasks, and successive variants have been proposed to further improve the performance
of pre-trained language models. In this paper, our goal is to revisit Chinese pre-trained language models to examine their effectiveness in non-English languages. This paper is based on the RoBERT model for fine-tuning, and experimental results show good performance on many NLP tasks.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Computer Application Technology submitted time 2024-01-06
Abstract： This report explores the application of the SGT model in the field of magnetic prospecting, with a special focus on its performance on the MGT, SNR0 and SNR5 datasets. The experimental results reveal that the SGT model suffers from high false alarm rate and large prediction bias when dealing with these datasets. To address the insufficient predictive and generalization abilities of the model, we designed a series of improvement experiments focusing on three aspects, namely, tuning parameter, optimizing the feature extraction method and modifying the continuity judgment.
Among these three improvement methods, tuning parameter achieved about 0.5% performance improvement, and the methods of feature extraction optimization and orthogonal basis judgment instead reduced the prediction effect by 20%. Through code review and logical reasoning, we found that the problem stems from feature extraction incompatibility with the model. In order to adapt to the orthogonal basis algorithm, we propose an improvement idea: introduce many different types of features, including time-domain features, frequency-domain features, and statistical features, etc., and comprehensively utilize the information of these features to construct a more complex and comprehensive SGT model. In addition, the stacking module is introduced to take the prediction results of a single model based on different features as inputs, and generate a more accurate ultimate prediction through further learning and synthesis.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Computer Application Technology submitted time 2024-01-06
Abstract： Spatial scene classification has long been a prominent area of research in the field of geographic information science. In the past, traditional approaches heavily relied on retrieval methods based on image features. However, given the rapid advancements in deep learning and artificial intelligence, the efficient classification of complex spatial scenes has become increasingly crucial. This paper presents a novel framework that combines object detection with knowledge graph to automate the process of spatial scene classification. Initially, the input images undergo processing using object detection techniques to identify key entities within the scenes. Subsequently, a knowledge graph, which encompasses various spatial scenes, entities, and their relationships, is utilized to identity spatial scene catogories. To validate the effectiveness of the framework, experiments were conducted using eight spatial scene categories as an example. The results demonstrated a high level of consistency with actual spatial types, thus affirming the efficacy of the framework and highlighting its potential application value in the domain of spatial scene classification.
Peer Review Status:Awaiting Review
Subjects: Information Science and Systems Science >> Basic Disciplines of Information Science and Systems Science submitted time 2024-01-06
Abstract： In recent years， the industrial process has been developing towards complexity and large-scale， which has posed a series of challenges for traditional fault diagnosis techniques to solve practical industrial process problems. With the superior performance and unique potential of deep learning in feature extraction and pattern recognition， the application of deep learning technology to fault diagnosis has become a current research focus. Therefore， this article introduces several typical fault diagnosis methods based on deep learning. Finally， the obstacles in the application of deep learning to fault diagnosis are discussed， and the future research directions are prospected.
Peer Review Status:Awaiting Review
Subjects: Information Science and Systems Science >> Control science and technology submitted time 2024-01-06
Abstract： One of the core topics of artificial intelligence is neural networks and deep learning, which imitate the working principle of the human brain and use multi-level neural connections to mine valuable knowledge and rules from data. The research of neural networks started in the 1940s and went through several ups and downs and innovations. It now covers many types and fields, such as convolutional neural networks, recurrent neural networks, speech recognition, computer vision and natural language processing. Deep learning refers to using multi-layer neural networks to solve complex nonlinear problems. It relies on massive data and computing resources, as well as efficient training and optimization techniques. Deep learning has achieved amazing progress in recent years, but also faces some difficulties and challenges, such as model interpretability, generalization ability, security and reliability. Deep learning is still a vibrant and promising research field, which is expected to open up more opportunities and possibilities for human intelligence and life. This article will briefly introduce some types of neural network structures and some deep learning model structures.
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Computer Application Technology submitted time 2024-01-06
Abstract： Diffusion models are a powerful type of generative model capable of producing high-quality results in various fields including images, text, and audio. This review aims to summarize and analyze the latest research progress in diffusion models applied in the vision domain, including both theoretical and practical contributions in the field. Initially, the article discusses the characteristics and principles of three mainstream models: denoising diffusion probabilistic models, score-based diffusion generative models, and diffusion generative models based on stochastic differential equations. It also analyzes derivatives aimed at optimizing internal algorithms and improving sampling efficiency. Furthermore, the review provides a comprehensive summary of current applications of diffusion models, including computer vision, natural language processing, time series analysis, multimodal research, and interdisciplinary fields. Finally, based on current trends and challenges, it offers a forecast for the future direction of diffusion models, aiming to guide and inspire research in the field. This article is intended to provide researchers with a comprehensive overview of diffusion model research and application, emphasizing its significant role and potential in the field of Artificial Intelligence Generated Content (AIGC).
Peer Review Status:Awaiting Review
Subjects: Computer Science >> Computer Application Technology Subjects: Medicine, Pharmacy >> Clinical Medicine submitted time 2024-01-06
Abstract： Pituitary microadenomas are usually difficult to detect by non-contrast MRI, and the risk of misdiagnosis is higher and the number of cases is small, which makes the detection, segmentation and classification of pituitary microadenomas difficult. Based on the above problems, a computer-aided diagnostic system DCEPM-CAD based on dynamic enhancement sequence is proposed. While extracting the dynamic enhancement MR sequence timing information, the attention module of HRNetv2 was added to the backbone network to improve. In order to avoid the problem that pituitary microadenomas occupy too few pixels in the image to extract their relevant features, this paper also introduces the TecoGAN image super-resolution method to super-resolution the pituitary region image. In a total of 862 MR image datasets of 275 eligible participants, the diagnostic accuracy of DCEPM-CAD for pituitary microadenomas reached 77%. At the same time, significant results were achieved in the segmentation of pituitary and pituitary microadenomas, and the similarity coefficients of Dice reached 92.16 and 72.54, respectively.
Peer Review Status:Awaiting Review
Subjects: Information Science and Systems Science >> Basic Disciplines of Information Science and Systems Science submitted time 2024-01-06
Abstract： Although the current 3D face reconstruction methods based on a single image can recover fine geometric details, these methods have limitations. The faces generated by some methods can't be really animated because they don't model how wrinkles change with expressions. Other methods are trained on high-quality facial scanning, and cannot be well extended to images of natural scenes. The method used in the report can return to the details of three-dimensional face shapes and animations, which are specific to individuals but can change with expressions. The model of this method can be trained to generate a UV displacement map from a low-dimensional potential representation composed of person-specific detail parameters and general expression parameters, while the regression quantity can be trained to predict details, shapes, expressions, postures and lighting parameters from a single image. In order to achieve this, this method introduces a new loss of detail consistency, which separates people-specific details from wrinkles that depend on expressions. This unwrapping makes it possible to synthesize realistic personal specific wrinkles by controlling expression parameters while keeping personal specific details unchanged. This method is learned from images of natural scenes, and there is no paired 3D data supervision.
Peer Review Status:Awaiting Review
Hits
Hits
Downloads
Comment