Cancer is a severe affection nowadays, leading to death in many situations. Among the liver tumors, the most often met is Hepatocellular Carcinoma (HCC), being present for 75% of the primary liver cancer patients. Despite its’ invasive character, the most reliable cancer diagnosis method is biopsy. For performing both non-invasive and accurate disease assessment, computerized techniques are required. Tumor semantic segmentation is useful in this context, achieving tumor detection, localization, and extension estimation at the same time. In our current research, we developed and comparatively assessed high performance methods for segmenting liver tumor, based on Convolutional Neural Networks (CNN), as UNet, UNet++ and DeepLabV3+, involving two CT image datasets in our experiments. The assessment was performed by considering both the polar and cartesian image representations. At the end, a maximum DICE value of 80.92%, a maximum IoU value of 69.55%, respectively a maximum accuracy value of 99.82% resulted.
Compared with teenagers in cities, unique groups such as teenagers living in mountainous areas have unique cultural experiences and language differences in using words to express emotions. This difference is reflected in aspects such as the use of local slang or dialects in emails. This article presents an improved BERT-based algorithmic model for detecting specific emotions in emails written by these teenagers and providing relevant feedback and emotional support. We classify adolescent emotions into: happiness, anger, surprise, sadness, fear and disgust. It was divided into training, validation and testing by analyzing the emails of 30 teenagers in the mountains. A combination of SVM, LSTM, and BERT algorithms was employed to enhance the emotion classification model's accuracy. The evaluation demonstrated that the BERT-based model notably outperformed others in both accuracy and F1-score.
In the domain of few-shot learning, where the scarcity of training data poses a significant challenge, this paper introduces an innovative approach. We present a few-shot classification algorithm that utilizes the Two-Stream Frequency Domain Information Network. This algorithm delves deep into the efficacy of training different frequency components of images in the context of few-shot classification. It integrates filtered and enhanced image frequency domain information as supplementary data to be adaptively fused with the original image data, enhancing and augmenting the available information. Experimental results establish that the few-shot classification model, employing the Two-Stream Frequency Domain Information Network, outperforms leading algorithms in the field of few-shot learning. This enhancement is evidenced by an accuracy improvement of 1.06% and 0.93% on the miniImageNet and tiredImageNet datasets, respectively, when compared to the Meta-Baseline, illustrating the efficacy of this specialized approach.
The key work of China’s high-speed railway has gradually shifted from "large-scale construction" to "large-scale operation", and the quality of high-speed railway operation has become the core focus in daily work. How to efficiently and reasonably evaluate the effectiveness of high-speed railway work has become a key research issue. This research focuses on the evaluation of train service plans. It analyzes the connotation of efficiency and effectiveness in train service plan evaluation and constructs an evaluation indices system from the following four perspectives: economic benefit, transportation efficiency, service effect, and passenger potential. Some key technologies including the management of data with complex logic and the display mode of different types of indices for the evaluation have been researched. Finally, the system has been developed and implemented based on the above research, which can provide intuitive and efficient auxiliary tools for evaluating and adjusting train service plans.
The aim of this research is to implement and evaluate a smart agriculture system based on big data analytics, which provides a new way to improve the efficiency of agricultural production in the context of glocalization and climate change. Our research utilizes advanced information technology and big data analytics to optimize the agricultural production process. The article begins by introducing the concept of smart agriculture and its importance in modern agriculture. Then, we describe the architecture of the smart agriculture system, including the key aspects of data collection, processing and analysis. The system collects farm data through a network of sensors and uses cloud computing technology for processing and analysis. For the evaluation of the system, we focus on both ease of use and performance. Through collaboration with agricultural practitioners, we collected feedback on the system's user interface, understanding of its functionality, and impact on the agricultural production process. In addition, system performance was assessed through indicators such as processing speed, data accuracy, and system stability. Test results show that the system performs well in terms of ease of use, with users being able to learn quickly and operate effectively. In terms of performance, the system demonstrated efficient data processing capability and stable performance. These results indicate that our smart agriculture system is not only technologically advanced but also user-friendly and can effectively assist in agricultural production decisions.
Mitigating the impact of imbalanced data on classifiers is a challenging task in machine learning. Oversampling, a well-known approach, modifies the data distribution and generates minority class samples to alleviate the influence of imbalanced data on classifiers. However, many existing oversampling methods often involve k-nearest neighbors (KNN) techniques, and changes in the KNN parameter values can lead to noticeable instability. The recently proposed Radial-Based Oversampling (RBO) addresses this issue, but it tends to introduce a significant amount of redundant samples and noise after sampling. In this study, we propose a Radial-Based Mixed Sampling method for imbalanced data classification. Firstly, we employ RBO to oversample a portion of the minority class samples. Then, we use the concept of potential to identify redundant samples and noise in the dataset. Finally, Radial-Based Undersampling (RBU) is employed as a data cleaning strategy to eliminate redundant samples and noise from the majority class samples. Extensive comparative experiments are conducted on 10 imbalanced datasets from KEEL, two different classifiers, and five sampling methods. The results indicate that our approach helps classifiers achieve higher F 1 scores and G-means. This suggests a significant improvement in the classifier’s ability to identify minority class samples while considering majority class samples. Additionally, statistical analysis shows higher Friedman Ranking for our method.
The ICP alignment algorithm requires high initial position of the point cloud to be aligned, and the alignment process is time-consuming, so a point cloud positional alignment algorithm based on improved nearest point iteration is proposed. For the error problem of mismatch removal, the adaptive threshold of mismatch removal is proposed, and its threshold is determined by the mean and standard deviation of the point distance set. To address the problem of underutilization of point cloud features by the cost function, a cost function based on the covariance matrix of the point distribution is established using the processing idea of probability theory. This cost function can also avoid nonlinear computation. The experiments show that this algorithm avoids the ICP alignment algorithm from falling into local optimum due to the initial position pose, and improves the alignment efficiency and accuracy at the same time.
The objective of this study is to propose an enabling strategy in complementary training competencies to generate qualified human capital in the mastery of advanced techniques in artificial neural networks. A horizontal case study is proposed which covers training stages in the student at an early stage for the training of a civil engineer in computer science. The case described here describes how the student, in the exploratory and training process, reaches a framework of maturity that allows him/her to validate specific competencies at different levels, such as applying computer science procedures, managing information technologies, software development, and modeling and applying engineering sciences. The coverage of competency integration courses allows maturity in knowledge, such as knowing, knowing how to know, knowing how to do, and being. The advances indicate a maturity of knowledge in the training process.
An integrated community energy system optimization method considering carbon emission reduction is proposed to improve its low-carbon operation level and reduce the total operating cost. Considering the uncertainty of wind power and solar power output, the uncertain power curve of new energy output is obtained by using scenario generation and scenario reduction methods. Ground source heat pumps are used to absorb excess renewable energy to decouple the "heat and power" problem. A carbon trading mechanism will be introduced into the integrated community energy system to further control carbon dioxide emissions and reduce carbon emission levels. Based on this, the optimal scheduling model of the integrated community energy system is established, which aims for minimizing the sum of daily operation cost and environmental management cost. Take a demonstration park as an example, and through setting different scenes for comparative analysis. The results show that the optimized scheduling method of integrated community energy system considering carbon emission reduction can reduce the cost and carbon dioxide emission of the system.
System logs record noteworthy information and become a valuable resource for tracking and investigating the status of a system. Detecting anomalies from logs as fast as possible can enhance quality of service. Although many deep learning algorithms have achieved good results in the field of log anomaly detection, we find that they assume that data is acquired and trained in bulk and may have high training time, while in reality log data streams are constantly arriving. Facing these challenges, this paper proposes a Long Short-Term Memory network (LSTM) based online log anomaly detection. We use transfer learning to train weak learners and integrate weak learners as strong learners for the constantly arriving log. We compare the performance of the methods using a public log dataset and experimentally validate the effectiveness of each module.
It is very common to have long tail distribution of items in recommender systems. However, due to the limited interaction records of long-tail items, recommending them to users significantly affects the model’s performance. In the existing methods, the features of the head item are transferred to the tail item, thereby ignoring the difference between them, and fail to produce a satisfactory recommendation effect. To address this issue, we propose a novel recommendation model based on meta-learning, called MLFRec. First, The MLFRec obtains the initial parameters from the head items through meta-learning, and then fine-tunes the model parameters in the learning process of the features of tail items; Second, it adopts attention graph convolutional network (AGCN) to enhance tail items, thereby reducing the difference between head and tail items. Through above two steps, the model takes advantage of the abundant data of the head items that help to address the issue of sparse data of the tail items, and improves the recommendation quality. Extensive experiments were conducted on three real-world datasets, and the results indicate that our proposed MLFRec model outperforms other state-of-the-art baselines for the tail items recommendation.
To optimize the motion planning of multi robotic arms for a single task execution, traditional methods rely on manually planning the end effector’s trajectory, which is time-consuming and labor-intensive. Reinforcement learning offers an alternative, where the robotic arm explores actions within its environment, receiving feedback through rewards, and learns control strategies for task completion. This research encompasses three primary areas: practical operation, simulation environment creation, and reinforcement learning algorithm training. For practical operations, the UR3 robotic arm model is programmed using a teaching device, facilitating motion planning via manual instruction. In simulation, a MuJoCo-based environment with a physics engine is developed to mirror the robotic arm’s real-world movements. This simulation controls the arm through a mocap data stream, guiding the end effector’s position and posture. Reinforcement learning training involves crafting various reward functions within the environment to achieve typical tasks, such as reaching targets and manipulating objects. Some preliminary results reveal that such a framework would facilitate the self-learning of multi agents in achieving the global target.
Geometric range query is an essential operation in many applications, including location-based service, querying on sensor networks, and computational geometry. In the Internet of Vehicles (IoVs) environment, the range query operations allow for retrieving vehicles or events within a specific spatial range, enabling applications such as traffic monitoring, incident detection, and task outsourcing. However, with the issues of latency, location and query range privacy, and geometric query range limits, the application of range query in IoVs still faces a number of serious challenges to its prosperity. In this paper, we design arbitrary hollow polygon preprocessing and points in polygon judgment strategies that support arbitrary hollow polygon range query. Furthermore, by integrating fog computing into IoVs, the low latency of the overall process is met. Finally, the theoretical analysis and experimental results show that the proposed scheme is secure and efficient.
With the increasing penetration rate of renewable energy, the balance cost of dealing with source-load uncertainty has attracted more attention in economic dispatching of power system. Especially in the electricity spot market, fully mobilizing the enthusiasm of flexible resources and reasonably sharing the balance cost have become the key challenges for system optimal dispatching and effective market organization. In this paper, considering the influence of source-load uncertainty on the reserve and ramping capacity demand of power system, a multi-period optimal dispatching model with chance constraints is established, and a market pricing method of flexible ramping service is proposed based on the moment definition deterministic equivalent transformation method and marginal cost pricing principle. Case study is carried out for PJM 5 node power systems, and the sensitivity analysis of violation probability parameters is carried out by Monte Carlo simulation. The results show that the above method can effectively describe the difference between energy cost and balance cost, and provide value measurement for flexible resources through detailed spatio-temporal price signals.
With the increasing popularity of network applications, internet information retrieval has developed into an important retrieval method for people to query and obtain information, providing a guarantee for users to obtain valuable information from massive amounts of information and improve the accuracy of information acquisition. In internet information retrieval, the application of data mining technology can significantly improve the efficiency of information retrieval. This article introduces data mining technology, analyzes its application value in internet information retrieval, and analyzes the specific application of data mining technology in internet information retrieval.
With the increasing complexity of industrial systems, new challenges are posed to the monitoring of industrial process data, which often appear to be characterized by nonlinear and strong feature correlation. Therefore, a sparse stacked denoise autoencoder(SSDAE) based anomaly detection model is proposed in the paper, which uses the autoencoder model to capture the nonlinear feature structure in the industrial data, and we extract the feature space and the residual space to build a statistic to capture the system changes, and finally use the KDE to determine the threshold to detect the anomalies. In this paper, using the Tennesse-Eastman dataset, method validation is carried out and compared with algorithms such as PCA, DAE and LRAE, which verifies the effectiveness of the algorithm, improves the detection rate of faults, and is able to identify more faults.
Foreign objects threaten transmission line reliability and can cause cascading blackouts. Manual inspection of thousands of miles is inadequate. This paper presents an optimized deep learning model for real-time foreign object detection. We enhance YOLOv4 with GhostNet, CBAM, and BiFPN to to simplify the model structure and improve efficiency and accuracy. GhostNet enables lightweight detection through efficient convolutions. CBAM and BiFPN refine representations. Experiments show state-of-the-art accuracy while using 4.4x fewer parameters than YOLOv4. Our model operates in real-time on edge devices, advancing automated inspection to monitor critical infrastructure. Precise and timely foreign object detection enables proactive maintenance to prevent faults and failures.
Deep neural networks have been widely applied across various domains, but their numerous parameters and high computational demands limit their practical usage scenarios. To address this issue, this paper introduces a convolutional neural network compression method based on multi-factor channel pruning. By integrating scaling and shifting factors from batch normalization layers, a multi-factor channel salience metric is proposed to measure channel importance. By removing redundant channels within the convolutional neural network, a compressed model is obtained. On the CIFAR-10 dataset, we pruned 93.06% of the parameters and 91.92% of the calculations from the VGG13BN network, with only a 2.81% decrease in accuracy. On the CIFAR-100 dataset, we pruned 72.84% of the parameters and 72.03% of the calculations from the VGG13BN network, with an accuracy improvement of 4.11%.
The prevalence of spinal tuberculosis (ST) is particularly high in underdeveloped regions with inadequate medical conditions. This not only leads to misdiagnosis and delays in treatment progress but also contributes to the continued transmission of tuberculosis bacteria, posing a risk to other individuals. Currently, CT imaging is extensively utilized in computer-aided diagnosis (CAD). The main features of ST on CT images include bone destruction, osteosclerosis, sequestration formation, and intervertebral disc damage. However, manual diagnosis by doctors may result in subjective judgments and misdiagnosis. Therefore, an accurate and objective method is needed for diagnosing of spinal tuberculosis. In this paper, we put forward an assistive diagnostic approach for spinal tuberculosis that is based on deep learning. The approach uses the Mask R-CNN model. Moreover, we modify the original model network by incorporating the ResPath and cbam* to improve the performance metrics, namely m A P small and F1-score. Meanwhile, other deep learning models such as Faster-RCNN and SSD were also compared. Experimental results demonstrate that the enhanced model can effectively identify spinal tuberculosis lesions, with an m A P small of 0.9175, surpassing the original model's 0.8340, and an F1-score of 0.9335, outperforming the original model's 0.8657.
Solving the under-determined system of temporally-dependent linear equations (UDSTDLE) problem has attracted extensive attention and study. With the fast developing of the artificial neural dynamics, some important and inspiring recurrent-class neural dynamics different from traditional numerical methods, such as Zhang dynamics (ZD), are proposed to obtain the solutions of UDSTDLE in a real-time manner. The ZD is also named ZNN (Zhang neural net). In this new paper, we combine the traditional gradient dynamics (GD) and the ZD to obtain the continuous gradient-Zhang dynamics (CGZD) model. Since digital hardwares and devices require discrete computational algorithms, we use five recently-new different TD (time-discretization) formulas, i.e., Euler forward (or termed forward-difference, or forward-difference-quotient) formula, Zhang-Taylor (ZT) discretization formula, six-instant (SI) discretization formula, eight-instant (EI) discretization formula, and ten-instant (TI) discretization formula of ZTD (Zhang TD [time discretization]) class, to effectively and efficiently discretize such a CGZD model so as to obtain five diffrerent discretization-type numerical-type algorithms. Finally, numerical simluations and experiments so as to further substantiate such correctness of such discretization-type numerical-type GZD algorithms.
