In this paper, an extended finite element method and a curve-fitting procedure are incorporated into an implicit level set method to construct a topology optimization approach inspired for manufacturing fields. The extended finite element method plays an essential role in dealing with a weak discontinuous model after utilizing a zero level set to be divided into the solid and void areas in each element. A Hamilton–Jacobi partial differential equation is transformed into ordinary differential equations over the entire design domain and updated by an upwind scheme. The proposed method is implemented in the minimum compliance design framework that has been extensively studied in topology optimization. Numerical examples illustrate the success of the presented approach in the accuracy, convergence speed and insensitivity to initial designs. It is suggested that the introduction of the nodal density variables interpolation with a level set and smoothed curve-fitting approach can be promising in material topology optimization.
International outbound call management presents significant challenges for multinational corporations due to the need to establish reliable and cost-efficient connections with numerous countries worldwide. The complexity of this task is heightened by the fact that different providers offer varying call prices and quality levels, making optimization and efficient management a daunting task for support teams. Efficient call routing is crucial for economic savings, avoiding poor call quality, and improving the overall Quality of Service (QoS) for users. In this paper, we propose a system that leverages data mining techniques to collect and analyze call logs primarily from the perspectives of costs and quality. The system is designed to develop a Smart Call Routing mechanism that optimizes routing decisions based on real-time data. With the consideration toward Multi-Objective Optimization (MOO) or Pareto optimization, we introduce a set of rules to evaluate and select the optimal routing path, focusing on maximizing cost efficiency while maintaining high call quality. This approach allows companies to optimize their call routing strategies, leading to significant cost savings, improved call quality, and proactive issue management. By minimizing potential business losses, our solution not only enhances operational efficiency but also contributes to a superior user experience.
Predicting stock prices is challenging due to the market’s inherent volatility and complexity. This paper explores how sequential models like Recurrent Neural Networks, Long Short-Term Memory, and Gated Recurrent Units can be used for stock price forecasting. To enhance prediction accuracy, Wavelet transform techniques are applied to clean the data from noise. Additionally, a dual-stage attention mechanism is introduced. This mechanism helps in capturing long-term dependencies and focusing on the most relevant data for prediction. In the initial stage, the input attention mechanism determines the most significant features at each time step by referencing the previous encoder hidden state. In the subsequent stage, the temporal attention mechanism identifies the most pertinent hidden states from across all time steps. This two-stage attention approach enhances both the accuracy of predictions and the interpretability of the model. The models are tested on various stock indices, demonstrating that combining data denoising with attention mechanisms significantly enhances forecasting accuracy.
A wireless sensor network (WSN) is a significant technological element of the Internet of Things (IoT). The core of a WSN consists of wireless sensor nodes (IoT nodes) that collect and transmit data from the installation sites to digital platforms. As IoT applications expand into various environments, the demand for low-power IoT nodes is becoming increasingly important. In this paper, we propose the schematic and implementation of low-power IoT nodes using power-saving techniques on the low-cost ESP8266 module. The IoT node contains the ESP8266 module and sensors to collect environmental data, including temperature and humidity then transmit via the ESP-NOW protocol. The ESP8266 module fully integrates both a Wi-Fi transceiver and a microcontroller, which contains central processing unit (CPU), system clock (SYC), and real-time clock (RTC) on a compact chip. However, in active mode, both the Wi-Fi transceiver and microcontroller are enabled, so the power consumption of the ESP8266 module is comparatively elevated, even when the Wi-Fi transceiver is in a state of waiting to receive or transmit data. To reduce power consumption, power-saving techniques of applying the sleep mode of the ESP8266 module was proposed in the design. Three sleep modes, including modem-sleep, light-sleep, and deep-sleep are provided by the ESP8266 module. Each sleep mode configures components such as the Wi-Fi transceiver, CPU, SYC, and RTC to be on or off differently. The IoT node was also programmed with the function of switching between sleep modes, from which users can select and configure the sleep mode and sleep duration accordingly. The power consumption of the IoT nodes is analyzed and evaluated between three sleep modes based on sleep duration and events. These IoT nodes are deployed in real-world environments to monitor various environmental conditions.
Facial Expression Recognition (FER) is a critical research area with wide-ranging applications. While advanced convolutional neural networks have enhanced FER performance, their complexity impedes deployment on resource-constrained devices. Inspired by the self-distillation architecture, this study introduces ResNet18-Lite, featuring a multi-branch architecture with Distillation Blocks designed in its shallow branchs. This novel design not only improves overall performance on FER tasks but also allows flexible branch selection during prediction, making it an optimal solution for resource-limited FER applications. Experimental results on popular FER datasets, including FER2013 and FER-Plus, demonstrate ResNet18-Lite’s superior performance compared to baseline deep learning models and other recent ResNet18 variants with Self-Distillation. Notably, ResNet18-Lite’s lightest classifier branch, utilizing only 0.9 million parameters, outperforms other more complex models.
In this paper, we present a system-on-chip (SoC) design for incorrect face mask-wearing detection on a low-cost FPGA device. The SoC architecture realizes a VGG9 model to classify the incorrect face mask-wearing persons. To accelerate the processing time, we proposed a convolution IP that can compute 32 convolutional calculations in parallel. The experimental result shows that our SoC system can process up to 15 image frames per second at an operating frequency of 130Mhz and the coverage score achieved up to 93.3% in real-time.
In the age of data-driven decision-making, generating synthetic data of high-quality to tackle issues related to data privacy, limited data availability, and the need for good datasets in machine learning. This paper looks at a few different GAN-based models for creating synthetic data. It focuses on their pros and cons when judged by different criteria, such as how useful they are for machine learning, how similar they are statistically, and how well they protect privacy. By conducting a comparative analysis, our research shows the potential advantages and drawbacks of each model. The findings indicate that, while GAN-based models are effective at producing realistic synthetic data, there remain challenges to solve in terms of scalability, variety, and ethical considerations. Potential areas for future research involve enhancing model scalability, developing hybrid models, and improving real-time data synthesis capabilities. This study emphasizes the importance of ongoing assessment and improvement of synthetic data generation methods to ensure their efficacy and ethical use in a wide variety of applications.
Medical records are one of the most sensitive types of data, so when applying machine learning models, it is necessary to ensure data privacy. In recent years, machine learning and pre-trained models have been developing rapidly, and medical data security when using those models is gaining strong appeal with researchers. This study proposes a federated learning model integrated with homomorphic encryption to enhance security and privacy while training machine learning models on medical datasets. Additionally, we conducted experiments with federated learning models using different data distribution ratios to evaluate the robustness of this approach. The results show that the CNN, ResNet50, ResNet152, and DenseNet169 models integrated with federated learning on the LIDC-IDRI dataset have comparable accuracy to centralized machine learning. Moreover, the federated learning model integrated with Homomorphic Encryption on the ResNet50 model showed a 4% increase in training time and a 36% increase in model size compared to federated learning without encryption.
In Van Hien University, which serves tens of thousands of students, efficiently addressing student inquiries related to training programs, tuition fees, graduation conditions, and output standards has become a critical and urgent challenge. Traditional advisory methods are increasingly strained by the volume of queries and the complexity of providing accurate, context-specific guidance. This research introduces an AI Agent leveraging the Advanced Retrieval-Augmented Generation (RAG) model to solve this problem. The AI Agent is designed to dynamically retrieve relevant information from extensive institutional documentation and provide precise, context-aware advice to students. By integrating cutting-edge language models with a robust retrieval mechanism, the proposed system aims to enhance the accuracy, relevance, and scalability of student advisory services. This approach not only improves the efficiency of responding to student needs but also ensures that the guidance provided aligns closely with the university’s standards and policies, ultimately contributing to a more streamlined and effective student support system.
In the ever-evolving financial market landscape, the detection and forecasting of anomalies play a pivotal role in investment decision-making and risk management. This study focuses on identifying anomalies within the VN30-Index of the Vietnamese stock market. The continuous fluctuations in stock data pose significant challenges for investors seeking to make accurate decisions. We address this issue by leveraging foundation models (FMs) in time series analysis, particularly utilizing the MOIRAI model based on the Transformer architecture. These models, pre-trained on large datasets, exhibit the capability to handle diverse and complex data efficiently. Our proposed methodology for detecting anomalies is based on probabilistic forecasting and comprises two main stages: fine-tuning the foundation model on VN30 data and calculating the deviations to identify anomalies. The results indicate that the fine-tuned model significantly improves forecast accuracy and successfully identifies critical anomalies in the VN30-Index. These anomalies highlight substantial trend shifts and can inform strategic investment decisions. Compared to traditional strategies like Bollinger Bands and Moving Average Crossover (MAC), our anomaly-based investment strategy demonstrates promising outcomes. This research underscores the potential of foundation models and probabilistic forecasting methods in supporting investment decisions, marking a novel approach for optimizing investment returns and risk management in stock markets.
In technology 4.0, all data is put on the Cloud for transmission and processing. Images are one of the most important data and are quite large because they are created every day. Therefore, the increased demand for image storage and transmission urges scientists to research image compression methods to reduce the excess image data and it can solve the above problem. This paper presents the Field Programmable Gate Array (FPGA) hardware architecture of the JPEG image compression method (the world’s most popular image compression method) using FDCT numerically calculated Floating Point combined with Pipeline. The main goal of the proposed method is to increase the processing speed and compression performance of the image. The tests were performed on Red-Green-Blue (RGB) standard images. Evaluated by execution time, mean error value (MSE), top noise ratio (PSNR), and compression ratio (CR). This complete design is synthesized on the Vitex7-VC709 board with a frequency of 100.7 Mhz, throughput of 250 FPS, and power consumption of 2.401 Watts.
Group testing is a longstanding problem with numerous applications in social life, it plays an important role in quickly performing sample tests and identifying and localizing the defect items on a large scale, in conditions where the number of tests is limited. The step of analyzing the pooling test matrix to predict the desired results requires certain technical probability calculations and combinatorial reasoning. In this study, we would like to propose an approach that uses a group of evolutionary algorithms to determine the patient’s situation based on information from performed tests and experiments with data samples on different materials to compare. Eventually, the paper evaluates the effectiveness of the aforementioned algorithms.
On E-commerce platforms, users not only rate items using numerical scoring systems, but also frequently share their opinions about item experiences through textual reviews. Multiple studies suggest that examining user reviews is an effective way to understand consumer tastes and identify key item characteristics, thereby enabling more effective recommendations. The core problems that this approach needs to address include effectively utilizing textual reviews for recommendations, specifically how to represent the reviews, identify essential information within them, and aggregate this information to obtain representations for both users and items. In this work, we introduce SBRec, a novel approach that employs pre-trained models to extract meaningful user and item profiles from textual reviews. SBRec incorporates two levels of attention networks to capture the hierarchical structure of information within a collection of reviews. At the lowest level, sentences within reviews are modeled using a pretrained transformer to capture the full context. The sentences in each review are combined through an attention mechanism to create a review representation. At the higher level, another attention mechanism aggregates information from all of a user’s reviews to generate their overall representation. The item’s representation is constructed in a similar way and compared with the user’s representation to estimate the rating. We experimentally evaluated the proposed method on five benchmark Amazon datasets. Our experimental results demonstrate that our method surpasses previous baselines and cutting-edge review-based recommendation techniques.
Machine learning for precipitation nowcasting using radar reflectivity images is a promising approach in weather forecasting. This paper proposes a new pipeline of using Convolutional Neural Networks (CNNs) integrated with Gated Recurrent Units (LSTM) and visual Attention mechanisms for quantitative precipitation nowcasting (QPN) using radar reflectivity images. We propose an end-to-end nowcasting network that leverages a Convolutional-LSTM with Axial Attention as the backbone to develop an adaptive forecasting system. Our results demonstrate the potential of deep learning techniques to significantly improve short-term weather predictions. The models were trained and validated using real-world Doppler radar data from Nha Be, Ho Chi Minh City, ensuring both practical relevance and applicability. Experimental results highlight that the combination of 3D Convolution, Convolutional-LSTM, and Axial Attention achieves superior performance, underscoring the effectiveness of this approach.
In recent years, IoT applications have become increasingly popular. Smart services have been deployed from the IoT infrastructure to provide convenience for humans in their lives and related activities. Alongside IoT’s potential, security and privacy concerns have been highlighted in the IoT architecture. One weakness of the IoT system is the widespread deployment of sensor nodes with wireless connections. Additionally, the limited resources of these sensor nodes pose a challenge in designing and implementing security solutions for the IoT infrastructure. In this article, we plan to deploy a Network Intrusion Detection System (NIDS) for IoT infrastructure. This system is towards to runs on the Swarm Learning framework, which supports decentralized machine learning models to ensure data distribution during training. This framework also operates on Ethereum – an open-source blockchain platform - to ensure authenticity and security while training decentralized machine learning models. We experiment with various scenarios using the DNN model via the CiCIoT2023 dataset and the CiCIoMT24 dataset. The results demonstrate that our proposed system ensures accuracy comparable to centralized machine learning and Federated Learning models. In addition, we also tested and evaluated based on training time and resource usage, thereby concluding that the cost and effectiveness of the Swarm Learning system is better than Federated Learning.
False predictions often hampered human action recognition in videos, reducing the reliability of detection models. This paper presents a novel approach that integrates Video Vision Transformer (ViViT) and YOLOv8 to minimize false alarms in action detection. YOLOv8 detects human subjects within video segments, while ViViT reclassifies these segments to reduce false positives. We validate our method on two benchmark datasets: THUMOS14 and EPIC-Kitchen. Our experiments substantially reduce false positives, improving model performance without sacrificing accuracy. Specifically, our framework reduces false predictions by 45.8This approach enhances the precision of action detection models, offering a more robust and reliable solution for practical applications such as video surveillance and human activity analysis in untrimmed videos.
The paper focuses on solving two problems. Firstly, we design a sensor station module of the parameters (CO,NO 2 , SO 2 ,O 3 , PM 2.5 , PM 10 ), which is real-time and can send data and store it on the IoT platform. The second is to predict environment parameters using deep learning networks for points that must be measured to predict future air quality. For the first problem, we consulted several open sources, and circulars on the AQI index to find the necessary air environment parameters and hardware suitable for the problem. However, there are many problems related to the sensor, specifically calibrating and finding the characteristic curve and evaluating the error of the characteristic graph for gas sensors (CO,NO 2 , SO 2 ,O 3 ), the IoT platform problem stores data using an available platform that supports APIs to store and display data. For the second problem, we have researched and consulted using the long short-term memory networks (LSTM) network model to predict air environment parameters. As a result, the system sent data within four months without errors. Besides, the parameter prediction part has relatively high accuracy up to 98.44% when deploying to predict future datasets.
Due to the increasing complexity of business processes and the growing number of required features, modern software rarely operates independently. Rather, it often relies on third-party services, many of which are available in the form of API calls, to perform recurring computing tasks. While software developers and programmers alike may enjoy the benefits of API calls (e.g., integrability, efficiency, shorter time-to-market), it is crucial for them to have a thorough understanding of API usage in the entire software code. Research in API usage analysis ranges from supporting the understanding of codebase semantics to identifying important software dependencies and recognizing threats from suspicious API calls. To our knowledge, most existing work in the realm is focused on the statistical properties of large codebases. While understanding these large codebases as a whole is important, gaining knowledge of how specific software utilizes APIs is equally crucial. In this work, we propose a tool proposal for visualizing API usage of a software project. To evaluate the effectiveness of our visualization techniques, we demonstrate using a couple of open-source programs in C#. The experimental results revealed some promising aspects as well as some issues that may be addressed in future research with regard to understanding API usage.
The local rank modulation scheme, a generalization of the rank modulation scheme, is proposed for accurate and effective data representation in flash memory. In this scheme, a sliding window traversing a sequence of real-valued variables yields a sequence of permutations. We focus on constant-weight Gray codes for the scheme. We represent two constructions for Gray codes of weights 2 and 3 proposed in [6] by grouping transition rules to create more easily calculable paths. From that we give calculations for the ranking and un-ranking on these Gray codes. Additionally, we show that the calculations are efficient by showing their running time in constant and logarithmic time for Gray codes of weights 2 and 3 respectively.
In the current digital transformation era and e-commerce, recommendation systems (RS) have become vital in optimizing profits and finding and reaching potential customers. Large companies (Facebook, Google, Amazon, etc.) possess huge amounts of data about customers and goods and it is important to exploit information to create customer suggestion systems. Suitable products will improve customer experience, save advertising costs, and sell more products. Therefore, it will make huge profits from marketing and e-commerce. This paper aims to build and optimize popular models for the problem of RS on the MovieLens dataset. Therefore, we test and compare the capabilities predictions of these types of models. In addition, this project uses PySpark technology to increase the amount of data aiming for execution on large datasets in practice to evaluate how the amount of data affects the ability to suggest. We also present the models and the construction installation steps environment and optimize their parameters. We assess each algorithm on the Movielens 100k and 1M. The results show that the model has MAE as 81.21% and 87.47% on two datasets Movielens 100k and 20M.
