Aili Wang, Fuling Li, Zhigang Cui, Shouhui Qi, Haisong Chen and Haibin Wu
This paper proposes YOLOv8-MultiSEAM, a multi-scale attention-augmented object detection framework tailored for X-ray security inspection scenarios. The MultiSEAM mechanism integrates multi-scale convolutional operations with channel attention mechanisms, employing dynamic feature weighting to precisely enhance critical details of dangerous items. This approach effectively addresses challenges such as small targets, occlusions, and illumination variations in complex logistics environments while maintaining computational efficiency. It significantly improves detection robustness and recognition accuracy under multi-angle imaging and high-noise conditions. Extensive experiments on two X-ray benchmark datasets, SIXray and HiXray, demonstrate that YOLOv8-MultiSEAM substantially mitigates multi-scale coexistence and high-density occlusion challenges compared to the baseline YOLOv8 model. The proposed methodology establishes a novel technical pathway for efficient dangerous item detection in complex security inspection scenarios.
Wenbai Liu, Haibin Wu, Zhigang Cui, Shouhui Qi, Haisong Chen and Aili Wang
With the widespread application of X-ray security technology in the field of public transportation, traditional manual security inspection methods have certain risks of false detection and missed detection, and more efficient and accurate automated solutions are urgently needed. This paper proposes a YOLOv11 improved model based on deep learning, combining space and channel collaborative attention mechanism (SCSA) to improve the accuracy and efficiency of hazardous goods detection in X-ray images. This model combines shared multisemantic spatial attention (SMSA) and progressive channel self-attention (PCSA) modules to enhance feature extraction and representation capabilities. The enhanced YOLOv11 achieves 92.4% mAP on SIXray (+1.1%) and 81.1% on PIDray (+1.9%), demonstrating consistent improvements over baseline. These results prove the effectiveness and universality of the proposed method, especially in complex backgrounds and small hazardous goods detection. In addition, the Heatmap analysis of the model further verified its good robustness.
Manman Yao, Haoran Lv, Mingji Yang, Aili Wang, Haibin Wu and Haisong Chen
The collaborative classification of multi-source remote sensing data is of great significance to improve the accuracy of ground object classification, but the modal differences and distribution inconsistencies between heterogeneous data make it difficult for traditional methods to fully explore the cross-modal correlation features. In this paper, a multi-modal joint classification framework based on heterogeneous double-branch network is proposed In order to meet the hierarchical expression requirements of LiDAR elevation features, a heterogeneous double-branch convolution network was designed: 1) two sets of parallel multi-scale convolutions were used to aggregate elevation information to enhance the representation of local details and global context; 2) Extracting Hyperspectral Imaging (HSI) spectral-spatial features and LiDAR elevation features, respectively. The twoway Cross-Attention fusion module was further constructed. Experiments on the Trento and Augsburg datasets show that the overall classification accuracy (OA) of the proposed method reaches 99.55% and 95.11%, respectively, and the classification performance of sensitive categories such as roads and buildings are significantly improved, which verify the superiority of the proposed framework.
Aissulu Kaldarova, Marco Vasquez, Nazym Baisbay, Alima Utemuratova and Madina Rakhimzhanova
Artificial intelligence (AI) is increasingly transforming English language education by offering innovative tools that support skill development. To examine the perceived benefits and limitations of AI in this context, a quantitative research design was employed using a structured survey method. A total of 137 university students from diverse academic backgrounds participated, providing valuable insights into the role of AI in their English language learning experiences. Data analysis revealed that AI tools—such as grammar correction software, automated feedback systems, and intelligent tutoring platforms—were positively associated with improved writing accuracy, vocabulary expansion, and learner independence. Nevertheless, respondents also expressed concerns regarding excessive reliance on AI, diminished opportunities for critical thinking, and reduced human interaction in learning environments. The statistical findings highlight both the pedagogical potential and the inherent challenges of integrating AI technologies in language instruction. These findings contribute to the growing body of research on educational technology and suggest that while AI holds significant promise for enhancing English language skills, thoughtful implementation is essential to maximize its effectiveness and minimize unintended drawbacks.
Davaajargal Myagmarsuren, Aili Wang and Haibin Wu
In this paper, we propose a novel deep learning architecture designed for multimodal multi-representation fusion framework that uses a new dual-path cross-attention fusion matrix to combine LiDAR structural information with complementary spectral and visual representations obtained from hyperspectral imaging. This approach combines an enhanced interconnected fusion matrix that uses self-attention and cross-attention processes to record intricate interactions across modalities with the potent feature extraction capabilities of ConvNeXt. Our approach tackles the problem of efficiently combining complementary data from several sensors while preserving spatial-spectral relationships. Our method surpasses the state-of-the-art multimodal fusion techniques, as shown by extensive trials on two benchmark remote sensing datasets. Our linked fusion matrix successfully captures intermodal relationships, resulting in a more accurate categorization of land cover, as demonstrated by the experimental findings. When compared to baseline methods), experimental validation across benchmark datasets demonstrates statistically substantial gains in classification accuracy, especially for classes that profit from the complimentary nature of spectral and visual characteristics (RGB).
Enkhbold Chuluunbaatar, Chantsaldulam Narandelger and Khos-Erdene Galbaatar
Aesthetic prostheses play a vital role in enhancing psychological well-being and social acceptance by replicating the natural appearance of lost body parts. This study applies “Full-Profile Conjoint Analysis” to investigate consumer preferences for aesthetic prostheses in Mongolia, a region with limited research on prosthetic adoption. Four key attributes were analyzed: realistic appearance, usability, material selection and cost. Using SPSS Orthogonal Design, 16 product variants were generated and evaluated. The results revealed that usability (27.372%) and material selection (27.558%) had the highest impact on consumer choices, followed by appearance (23.942%) and price (19.855%). These findings provide valuable insights for improving prosthetic design, manufacturing, and affordability in Mongolia, laying the groundwork for future research and development in the field.
Mijiddorj Khayankhyarvaa, Enkhbold Chuluunbaatar and Cong Guan
It presents a comparative analysis between traditional and artificial intelligence (AI)-driven design methodologies applied to Mongolian traditional clothing. The objective is to evaluate the effectiveness of AI in enhancing design efficiency, quality, and cultural preservation. Through empirical case studies and comparative trials, the research demonstrates that AI technologies, such as generative adversarial networks (GANs) and 3D simulation tools, significantly shorten design cycles, improve output fidelity, and facilitate the digital archiving of cultural heritage. The findings highlight the potential of AI to revolutionize the fashion design industry while preserving the cultural significance of traditional garments.
Yalin Li, Aili Wang and Haibin Wu
Hyperspectral imagery (HSI) classification faces challenges stemming from intricate spectral signatures and distributional discrepancies. In this paper, we introduce a new classification method that fuses spatial-spectral attention and domain generalization technology, and propose a hybrid deep attention architecture that integrates CNN and transformer for dual-branch (local-global) hyperspectral classification, enabling efficient extraction of discriminative spectral features. In addition, we use a coordinated optimization framework with a Lion optimizer and cosine annealing strategy to reduce local optimality. Results from experiments on various HSI datasets indicate the proposed method surpasses state-of-the-art HSI classification models, which validates its efficacy in cross scene HSI classification.
Feruza Zakirova, Madina Zakirova and Shakhnoza Pozilova
This study investigates the influence of demographic factors such as gender, academic rank, region of residence, and field of specialization on the participation of third-age information and communication technology faculty members in the distance professional development in Uzbekistan. Using data of 157 participants from the national portal, the analysis identifies key demographic trends in the participation forms, including the alternative program and distance training, as well as unregistered participants and did not completers. The results indicate statistically significant associations between specialization and participation forms, while gender, regional, and positional disparities highlight the need for more targeted strategies in professional development planning. The findings contribute to developing effective adult learning programs, aging, technology, and continuous professional development research.
Dilnoza Zaripova, Nasibakhon Rasulova, Golibjon Otamurodov and Durdona Primova
This study presents the methodological experience of implementing an AI-focused educational project conducted by first-year master’s students at the Tashkent University of Information Technologies. The initiative was structured using an integrated educational model combining CDIO (Conceive – Design – Implement – Operate) and PBL (Project-Based Learning), which ensured a balance between systematic engineering design and student-centered problem solving. Throughout the project, students engaged in all essential phases of digital product development — from problem identification and research to implementation and presentation. The project involved the use of tools such as Python, Whisper speech recognition, and various open-source libraries for audio and text processing. As a result, students developed both technical competencies — including programming, speech recognition, and debugging — and soft skills such as teamwork, project planning, time management, and oral communication. The paper discusses implementation challenges related to tool integration and varied student preparedness, alongside strategies for instructional support. It also presents comparisons with international CDIO and PBL practices in AI education, emphasizing similar approaches in global institutions. The outcomes demonstrate the model’s effectiveness and its potential for application in engineering and teacher education programs.
Shadi Saleh, Ganpat Bahadurlal Kalal, Batbayar Battseren and Wolfram Hardt
High-quality datasets are critical for ensuring the reliability and trustworthiness of AI models, as inconsistencies in training data can significantly degrade performance. This paper introduces a scientifically rigorous, cloud-based framework for automated dataset validity analysis and enhancement, designed to address key data quality issues such as bounding-box inconsistencies, annotation errors, and semantic inaccuracies. The framework, integrated into the SmartCityCloud environment, employs automated validation and correction procedures to systematically improve dataset reliability. A case study using an “Insulator” image dataset and the YOLOv8 object detection model demonstrates the effectiveness of the proposed approach. Experimental results show that these enhancements significantly improve model performance metrics, including precision, recall, and mean average precision (mAP), with observed mAP gains of up to 16.1% compared to baseline methods. These findings underscore the transformative impact of robust data preprocessing on AI model accuracy and reliability. By providing a scalable, real-time, and automated solution for dataset monitoring and improvement, this framework establishes a foundation for more trustworthy AI-driven decision-making systems and paves the way for advancements in automated data quality assurance.
Buyandelger Banzragch, Lut-Ochir Vanganjil, Oyungoo Enkhbold and Munkhmagnai Altangerel
This study examines the critical role of developing citizens’ information and communication technology (ICT) competencies in promoting participation, productivity, and equity in the knowledge-based society of the 21st century. Drawing on case studies and policy analyses, the paper identifies effective programs and opportunities for enhancing fundamental ICT skills. Strengthening citizens’ digital competencies is shown to be a key factor not only for personal growth but also for supporting the sustainable development of society. The study further underscores the importance of fostering lifelong learning to enhance personal development, employment, social participation, access to digital education, and the formation of digitally capable citizens.
Mangaljalav Chimed, Munkhtuya Erdenebat, Chimed Orshuu and Batgerel Tumurbaatar
Mongolia’s power system is currently underdeveloped and underutilized, with consumer capacity last year limited to around 150 MW to manage peak winter loads. The current and future energy sector policies of the Mongolian government emphasize in-creasing the share of renewable energy, particularly through the installation of high-capacity wind power plants in the Gobi region. Efficient and optimal utilization of large-scale renewable energy resources and their integration into the national energy system have become critical challenges for Mongolia. In this study, we analyzed the wind energy resources in the Southern Gobi region of Mongolia, based on wind resource measurements, followed by mathematical processing and resource evaluation. We determined the potential capacity for wind power plants based on these resources. Additionally, we examined the impact of integrating large-scale wind power plants into the southern region’s power grid on parallel operation modes. The wind resource assessment was conducted using Windographer simulation software.
Zagdkhorol Bayasgalan, Enerelt Bayaraa, Davaakhuu Battulga and Tsetsgee Bayasgalan
Frequency is inextricably linked to the reliability and stability of power systems. The purpose of this study is to clarify the causes and consequences of frequency changes, which are common abnormal conditions in power systems, and to study the theoretical basis of frequency control controls and adjustment methods related to these controls. Using control simulation software, they artificially create real problems in power systems and then simulate how the frequency changes and the stabilization process with the help of controls, discuss the results and conclude. Therefore, this study artificially creates an overload on a load, controls the frequency of the generator feeding the load, and then controls how it decreases and the frequency stabilization process is tested in detail through simulation.
Jargalmaa Gombo-Ochir and Altantovch Ganbaatar
Artificial Intelligence (AI) is reshaping online education by enabling personalized and adaptive learning systems. This study introduces a modular AI-powered learning framework that dynamically tailors instructional content and pacing based on real-time learner data. The system integrates a Gradient Boosting Classifier, SHAP-based explainability, and federated learning for privacy-preserving analysis. Evaluation on a synthetic dataset reveals modest model accuracy but offers insights into learner behavior patterns. Limitations—such as the use of simulated data, lack of real-world testing, and no pedagogical user validation—are acknowledged. Ethical considerations, including privacy, fairness, and data simulation protocols, are discussed. This work outlines a foundation for future deployment of AI-enhanced, learner-centric online education environments.
Chen Wang, Narantsatsral Delgerkhuu and Lin Hu
This paper focuses on the study of image-based virtual and real restoration technology in the era of artificial intelligence, aiming to solve the problem of restoration of ancient murals damaged by natural and human factors in Mongolia’s Shoroon Bumbagar Tomb. AI technologies-deep learning, generative adversarial networks for example-can analyze mural image, and accurately restore damaged areas. The findings demonstrate that this method outperforms traditional approaches in mural restoration, enhancing both accuracy and efficiency while offering a novel technical solution for cultural heritage preservation.
Figan Yaftali, Reda Harradi and Wolfram Hardt
Multi-laser scanning sensors, commonly known as Light Detection and Ranging sensors, are increasingly used in autonomous unmanned aerial vehicles to enhance real-time environmental perception and modeling. As demand for unmanned aerial vehicles grows, reliable and safe landing technologies become essential. Light detection and raging sensors’ precision and efficiency make them valuable across various fields like robotics, agriculture or disaster management. A robust system using the Benewake TFmini Plus sensor and Arduino Uno was developed for real-time obstacle detection and avoidance. A simulation environment employing the Potential Field Algorithm further demonstrated dynamic obstacle avoidance and safe zone identification. Results highlight the role of Light Detection and Ranging sensors in enabling costeffective, scalable unmanned aerial vehicles’ navigation and safe landing.
Hanli Xu, Zi Wang, Yangyang Liu, Yihong Zou, Wenlan Wang, Ruming Song, Hui Cao, Hongwen Li and Xiuping Hanv
This study investigates gender differences in daytime nap duration using a large-scale, objective, and longitudinal dataset collected via AI-enabled smart beds. Data from 3,272 participants across 200 Chinese cities were recorded continuously over three years (2021–2024), generating over 160,000 nap records. Results showed that women generally had longer nap duration than men and that nap behavior exhibited a non-linear trend over time. Model comparisons revealed that a cubic regression model best captured the dynamic patterns in nap duration. These findings contribute to the growing literature on sleep behavior by highlighting the importance of gender, time, and methodology. This study demonstrates the value of AI-based monitoring in enhancing ecological validity and provides a foundation for future research on the biological and sociocultural mechanisms underlying nap behavior.
Odgerel Ayurzana
This paper presents the design, development, and implementation of a voice-controlled robotic arm featuring three joints and five degrees of freedom. The system integrates a smartphone application, developed using MIT App Inventor, with a microcontroller to enable wireless control via Wi-Fi. The application utilizes Google’s cloud-based Speech-to-Text API to convert spoken commands into text, which is then processed and transmitted to the robotic arm. The arm comprises six servo motors and multiple joints, enabling precise multi-directional movements and object manipulation. Inverse kinematics was employed to determine the rotational angles of the robotic arm’s joints. It is capable of executing commands such as rotating left or right, moving up or down, gripping, and releasing objects. A usability study involving six participants demonstrated an average command recognition accuracy of 89%, with performance influenced by the clarity of English pronunciation. These results indicate the system’s potential for accessible, lowcost robotic control in various applications.
Batshagai Baatar, Bayar-Erdene Lkhagvasuren and Uranchimeg Tudevdavga
This paper describes the design, implementation, and empirical evaluation of a 3D virtual reality (VR) laboratory tailored for dental education and its evaluation based on students’ perceptions. Practical lessons for dentist students are essential important to learn correct movement of fingers and hands with special dentist tools and equipment. In reality to have live practical lessons are highly limited. Therefore, immersive simulation of tooth extraction procedures using Meta Quest 3 headsets and a custom-designed interactive environment-based 3D laboratory comes in teaching and learning in the classroom. The developed laboratory as a pilot test used for teaching at Etugen University in winter semester 2024/2025. In total 51 students are enrolled in that course and applied the 3D laboratory in learning. With target to figure out acceptance and satisfaction of students relating to offered laboratory the corresponding survey was designed sent to students. The result of the survey shows that 78% of respondents agreed that integrating 3D laboratories into education is appropriate, while 74% rated it as efficient. Notably, 64% confirmed that this improved understanding contributed to better learning outcomes in the course. This is the only beginning of complete VR laboratory for dentist students’ practical lessons and authors improves the 3D environments based on the students’ feedback.
Munkhtuya Uurtsaikh, Magvan-Erdene Gantulga and Dolgorsuren Batjargal
In Mongolia’s healthcare sector, medical prescriptions are predominantly handwritten by physicians and directly handed to patients. This traditional practice poses several challenges, including increased workload for healthcare professionals, risk of lost or damaged prescriptions, and limited access to essential medications. Moreover, paper-based processes contribute to negative environmental and economic impacts. To address these issues, this study proposes the development of an AI-powered smart prescription system that automates the prescription process using advanced artificial intelligence technologies. The proposed system captures and processes doctors’ spoken instructions during patient examinations and generates structured prescriptions in accordance with the MNS 5376:2016 national standard. These prescriptions include details such as medication names, dosages, preparation methods, and usage instructions. Additionally, the system is designed to integrate with pharmacy information systems and the E-Mongolia government platform, thereby streamlining communication among patients, pharmacies, and healthcare providers. Unlike e-prescription systems developed in high-resource contexts such as the United States, the United Kingdom, and Sweden, the proposed solution is tailored specifically to the Mongolian context—a low-resource language environment with distinct linguistic and infrastructural challenges. The research also examines comparable systems implemented internationally and incorporates relevant best practices to enhance the effectiveness and contextual suitability of the proposed system.
Tseren-Ochir Purevjargal and Kojiro Watanabe
Understanding spatial patterns of urban population change is critical for effective urban planning, especially in regions undergoing both growth and decline. This study investigates demographic change at the 500 m mesh level in Tokushima City, Japan, between 2010 and 2020, using a spatial machine learning framework. A Random Forest (RF) model was trained with 2010 population data and 24 spatial features, such as zoning ratios, elevation, disaster risk, and proximity to infrastructure to classify population change (increase, decrease, no change) and estimate its magnitude. The classification model achieved accuracies of 67.91% in Urban Promotion Areas (UPA) and 74.91% in Urban Control Areas (UCA), with key features including pop2010, building density, and access to clinics. This study demonstrates the utility of fine-grained, mesh-based analysis in capturing localized demographic dynamics and supporting data-driven urban planning in shrinking cities.
Batbayar Battseren, Shadi Saleh and Wolfram Hardt
This paper presents a lightweight geolocation method for MAV-based forest inspection using only monocular RGB video and onboard sensor data (GPS, yaw, pitch, roll). Unlike existing approaches that rely on stereo vision or depth sensors, the proposed system operates entirely in postprocessing without requiring any additional hardware. We introduce a triangulation-based technique that projects detections—typically tree crowns—onto the ground from multiple MAV viewpoints, forming a “cast-shadow” effect. By intersecting these projections, the system estimates true object locations with high accuracy. A two-stage tracking and clustering process further refines these estimates to yield one coordinate per tree. The system achieves 2.5–3.0 m average geolocation error in real-world tests, independent of object height or viewing angle. The approach enables scalable, lowcost, and sensor-free forest geolocation using standard MAV footage.
Uranchimeg Tudevdavga, Battsetseg Erdenebat, Paul Hubert Vossen, Nikolay Lomakin and Ilkhom Boynazarov
In this paper we will describe the main concepts, contents and user requirements regarding the Artificial Intelligence powered Assistant for Lecturers (AIpA4L). It will be based on our own ideas and a survey among future users of such an assistant. The main idea for AIpA4L arose due to observed weaknesses or omissions of recently proposed AI- tools for use in daily life of many people. Chatbots or other smart assistant systems say hello to customers of on-line banking, of citizen services and of many other on-line services. However, such smart services are still underdeveloped in educational institutions. The proposed AIpA4L is meant to fill this gap. The main task of AIpA4L can be described as a 24/7 available up-todate communication service for students, lecturers and other staff. One highlight of AIpA4L – not yet implemented in earlier versions of AI-based assistants for lectures - will be a broadly conceived taxonomy of topics from which the user will select the one s/he is currently most interested in. We have run a checklistbased survey among students and lecturers to collect a preliminary overview of topics of interest. In total, 117 students of MUST were invited to respond over a period of two weeks. The feedback was quite high: almost 80%. The respondents selected required topics related to teaching and learning. The result of the open-ended question “What do you primarily expect from an AI powered assistant?” confirmed the result from the “required topics” closed questions. The main concepts, contents and user expectations regarding the AIpA4L service will be described and discussed in this paper.
Mikhail Khnyunin, Mikhail Grif and Maxim Bakaev
Machine vision techniques are increasingly used to facilitate communication for the deaf people and make them more socially involved. In this paper, we describe the development of a Russian dynamic sign language translation system based on a hybrid approach combining two solutions. The first is the long short-term memory (LSTM) network with a component approach that determines the configurations of a signer’s hands and their spatial position. The second is the usage of a convolutional neural network (CNN) to analyze the images obtained by merging the successive frames with a given pixel lifetime. The output of the two approaches is combined using a custom dictionary of subclasses based on our dedicated dataset. Our solution is based on the MediaPipe framework, which allows obtaining three-dimensional coordinates of key points of hands and the human body. This can be done using a regular camera, so the proposed solution can be used ubiquitously. We were able to achieve 88.2% accuracy with a recognition time of one fragment of 50 frames in 0.57 seconds, which can allow real-time sign language recognition.
Djamshid Sultanov, Khusniya Akhmedova and Hyunchul Ahn
Measuring semantic similarity between sentences is central to plagiarism detection, authorship attribution, question- answering, and many other NLP applications. We consolidate existing Vector-Space and deep neural approaches into a single text-processing pipeline (TPP) for low-resource Uzbek. The TPP supports flexible tokenization, optional emoji handling, and alternative embedding stages (TF–IDF, Word2Vec, Doc2Vec). A lightweight API layer feeds sentence embeddings to a modular paraphrase-classification component that hosts distance-based or neural models. Our design is validated on publicly available Uzbek corpora and synthetic sentence pairs; cosine similarity combined with a logistic-regression meta-classifier yields up to 0.86 F-score without external WordNet resources. The pipeline is released as an install-and-play prototype for educational use.
Selenge Munkhbayar, Stephen Karungaru, Kenji Terada, Otgonbayar Bataa and Altangerel Ayush
With the rapid advancement of technology, integrating intelligent systems into agriculture has become a significant area of innovation. Traditional tomato harvesting is laborintensive and time-consuming, prompting the need for automation. This study focuses on accurately classifying tomato ripeness to support automated harvesting. Specifically, we employ deep learning techniques to categorize tomatoes into three ripeness levels based on color and identify those ready for harvest. We evaluated object detection models, particularly the You Only Look Once (YOLO) family, for their effectiveness in detecting tomato ripeness. The best-performing model achieved a mean Average Precision (mAP@0.5) of 0.898, or 89.8%, indicating strong potential for practical application in robotic harvesting systems.
Khulan Khalzaa, Stephen Karungaru, Kenji Terada and Tsend-Ayush Chimed-Ochir
Vision-language models (VLMs) have demonstrated strong capabilities in image captioning, yet their reliability in structured, domain-specific environments such as traffic scenes remains underexplored. In this work, we introduce a five-level synthetic dataset of urban traffic scenarios, designed to represent progressively increasing scene complexity, from static environments to anomalous and high-interaction situations. To address the grounding limitations of existing captions, we define grounded objects as visually detected entities (e.g., vehicles, pedestrians) that should be explicitly referenced in textual descriptions, ensuring alignment between visual content and language.
We propose a model-agnostic refinement module that filters hallucinated content and inserts factual references to such omitted but visually grounded objects. We demonstrate the utility of this approach by applying it to Moondream2, a lightweight VLM, as an initial use case, showing that the refined captions better align with scene content and improve interpretability under diverse-complexity conditions. This lightweight refinement strategy highlights a practical pathway toward enhancing factual consistency in captioning systems, motivating future work on relational reasoning and risk-aware scene understanding.
Munkhbaatar Dagvadorj, Tsedevsuren Danzan and Purevdolgor Luvsantseren
The primary objective of this study was to identify the level of digital competence among university lecturers based on student evaluations and to group lecturers with similar evaluation patterns using clustering analysis. A total of 514 undergraduate students from seven public and private higher education institutions in Mongolia—namely, Mongolian university of science and technology (MUST), Mongolian National University of Medical Sciences (MNUMS), Mongolian State University of Education (MSUE), ETUGEN, IDER, and ACH Universities—participated in the study. The data, comprising responses to seven evaluation items, were analyzed using K-Means and K-Medoids (PAM algorithm) clustering methods. Key statistical indicators such as Cronbach’s alpha (≥0.947), the Silhouette coefficient (≥0.88), and the Hopkins statistic (≥0.965) confirmed the reliability of the questionnaire and the suitability of the data for clustering. Both clustering techniques effectively categorized lecturers into two distinct groups labeled as “Average” and “Good,” with average ratings ranging from 2.2–2.6 and 3.9–4.4, respectively. The K-Medoids method, being medoid-based and more robust to outliers, produced more reliable results. Between 36.2% and 39.1% of students were grouped under the “Average” category, while 60.9% to 63.8% fell under the “Good” category, indicating a strong potential for differentiating lecturers’ digital competencies. One-way ANOVA revealed statistically significant differences (F > 300, p < 0.001) between the two clusters for each of the seven evaluation items (A1–A7), confirming the divergence in student assessment patterns. This study demonstrates that student feedback can serve as a credible basis for assessing university lecturers’ digital competence through robust, data-driven methods.
Batdorj Davaagombo, Namnan Tumurpurev, Tsolmonbaatar Danjkhuu and Ulziisaikhan Purevsuren
This paper describes to study the design of a device that cleans fine wool and cashmere fibers falling into the lower zone of the receiving wheel of the carding machine. The aim was to determine the boundary layer of the airflow formed on the working surface of a high-speed cutting wheel and to study the effect of unsteady flow formed between the teeth of the saw. Theoretically determine the movement of wool fibers detached from the toothed surface of the drum in the boundary layer and experimentally determine the zone where fine wool fibers fall in large quantities. When the machine is operating smoothly, an average of 8-10 g of wool falls every 5 minutes on an area of 0.12 m2. The distribution of wool fibers falling in the lower zone of the cutting wheel was divided into zones about 70% of the falling cashmere particles fell in zone 3. The most important issue is whether or not changing the design and some operating parameters of combing machines and equipment can have a positive effect on the quality of the final product.
Munkhjargal Nergui and Ganbat Danaa
This study explores the integration of the WorldSkills Mechanical Engineering-CAD professional standard into the CAD-II course curriculum within the Technical Vocational Education Program for Industrial Equipment Maintenance and Operation. The training process was redesigned to align with this standard, and its effectiveness was evaluated through the results of the Mongol Skills-2025 school-level competition and the CAD-II course assessments. Findings from these evaluations were used to inform and plan future research initiatives aimed at further enhancing the curriculum and student outcomes.
Ayurzana Badarch, Boldbaatar Nyamjav, Nasanbayar Narantsogt, Uranzaya Bayaraa, Batzorig Gansukh and Gantulga Tserendorj
Hydraulic engineering education is rapidly evolving to meet the dual challenges of climate change and digital transformation. This paper presents the comprehensive reform of the Hydraulic Engineering Master’s Program at the Mongolian University of Science and Technology (MUST) between 2021 to 2024, with a focus on climate change adapted and climate-resilient water management and the integration of hydroinformatics. Transforming Hydraulic Engineering master program toward climate change adaptation and digitalization is essential to equip graduates with advanced skills for resilient water management, sustainable infrastructure, and innovative decision-making in a rapidly changing environmental and technological landscape. Supported by the Erasmus+ CCWater project, the program introduced an innovative curriculum featuring big data analytics, integrated water resources management (IWRM) in changing climate, and storm water management with low impact development. The transformation was driven by strategic stakeholder engagement, international partnerships, pedagogical innovation, and significant laboratory investments. The outcomes—improved educational quality, research capacity, and global visibility—offer a model for modernizing engineering education in similar contexts, ensuring relevance in the face of environmental, technological, and labor market shifts.
Battamir Adilbish, Ankhbayar Nyamdavaa, Kiran Kaladharan, Gordon Cichon, Fan-Gang Tseng and Tseren-Onolt Ishdorj
This study develops a machine learning approach to detect SARS-CoV-2 variants (Beta, Gamma, Omicron) using surface-enhanced Raman spectroscopy with a barcode-based multiplex assay. A dataset of 54 SERS spectra, comprising 28 Positive and 26 Negative samples, was analyzed, targeting characteristic peak intensities at 380 cm−1 (Omicron), 540 cm−1 (Beta), and 1336 cm−1 (Gamma). Raw spectra were preprocessed with baseline correction, normalization, and smoothing to minimize noise. Features, including peak intensities, their ratios, and total spectral intensity, were selected to distinguish Positive and Negative samples. Three machine learning models SVM, RF, and LR were trained on 80% of the data (43 samples) and tested on 20% (11 samples), using five-fold cross-validation to prevent overfitting. SVM achieved the highest accuracy (90.00%) and perfect precision (1.0000), followed by RF (81.82%), while LR recorded the lowest accuracy (72.73%), limited by the small dataset and complex, non-linear SERS spectra. The framework supports extension to other pathogens, such as Influenza A/B, using additional reporters. The limited dataset size necessitates further data collection and validation. Future work will explore deep learning models, like convolutional neural networks, and employ Gauss-Lorentz methods for data augmentation to enhance dataset size and model robustness for automated, pointof- care diagnostics.
Battsetseg Bayarsaikhan, Nasanbayar Baavgai and Jadamba Badrakh
This study aims to evaluate the self-assessed competencies of biomedical engineering graduates with 0–3 years of work experience working in Mongolia’s health technology sector, alongside assessments provided by their employers. Assessing graduate competencies is a vital tool for determining educational outcomes, enhancing the quality of education, and supporting post-graduation professional development. The research utilized a questionnaire comprising 45 items across 19 competency domains, developed based on international frameworks such as CDIO [1], ABET [2], and OECD standards, as well as national guidelines. Data were collected from 42 graduates across three universities. Although the sample size is limited, the findings provide important insights that can serve as a foundation for improving educational programs. The results reveal varying levels of confidence in professional, communication, and technical skills among graduates, indicating the need to strengthen postgraduate training and practical education.
Jonathan Sande, Bilguun Erdenebaatar, Tseren-Onolt Ishdorj, Chuluuntsetseg Damiran and Tuyatsetseg Badarch
Access to high quality lexical datasets such as full-language word lists and lexicons is essential for the development of software applications like input method editors, spell checkers, and dictionaries. Although such resources may exist for a given language, permission to use them is sometimes hindered by copyright licenses. This is especially true for low-resource languages. This study focuses on three methods for using large language models to generate extensive lists of accurately spelled Mongolian words, a low-resource language, by comparing direct list generation, extraction from topically generated texts, and translation of a comprehensive English word list. Direct list generation was ineffective (producing a list of only 4,286 words with a 3.5% error rate). List translation was able to generate 23,176 unique words but also suffered from a high error rate (7.5%). Article extraction was more promising, producing a list of 15,234 words with a low error rate of 0.8%.
Delgermurun Batbold, Zagarsuren Batbaatar, Purevtseren Bayarsaikhan and Uuganbayar Purevdorj
This paper proposes a design of toroidal coil energy harverster, it harvests electrical energy from the radio frequency (RF) energy of the leaky feeder communication system. The leaky feeder system in UHF band is used as a main telecommunication method in the underground tunnels. Many kinds of sensors for the environment monitoring in the tunnels have been worked, that requires low power energy source. In addition, special needs of this kind of design can be mentioned derived from real problems, which can be low power LED blink cable identifier, RF power level meter for dusty and dusk environment of tunnels of underground mine. The simulated results of proposed design show power capacity of 0.754 W, it is sufficient to energy demands of low power devices that require 0.5 to 5 mW of power. This approach offers alternative potentials such as self-sustaining and maintenance-free compared to conventional battery-powered systems.
Erdenetuya Erdenebileg, Uuganbayar Purevdorj, Purevtseren Bayarsaikhan and Erdenebayar Lamjav
Efficient wireless network planning in indoor environments remains a critical challenge due to complex spatial layouts, material-induced attenuation, and dynamic coverage requirements. This paper proposes a novel AI-assisted framework that combines deep learning with automated optimization to predict wireless signal propagation and strategically place access points (APs) with minimal blind zones. A convolutional neural network based on SDU-Net architecture is trained using synthetic floorplan datasets that incorporate material properties and randomized AP configurations. The model predicts coverage heatmaps, which are then used to evaluate and select optimal multi-AP placements through blind spot minimization. Simulation results demonstrate that the proposed method significantly improves coverage uniformity and reduces planning time compared to traditional manual and rule-based strategies. This approach provides a scalable, data-driven solution for indoor wireless deployment scenarios, with potential applications in smart buildings, industrial networks, and next-generation communication systems.
Bolormaa Ayurzana and Zolzaya Choijin
This study investigates the reliability, validity, and user perceptions of the EnglishScore test as a digital alternative for academic English assessment among students at Mongolian public universities. Aligned with the Common European Framework of Reference for Languages (CEFR) and motivated by national education reforms, the EnglishScore test offers a mobile-based platform designed to assess core English skills efficiently. Data collected through a structured questionnaire reveal that while many students appreciate the test’s accessibility, convenience, and quick results, concerns persist regarding its comprehensiveness and comparability to established exams such as IELTS and TOEFL. Technical challenges, including unstable internet connections and time constraints, also affect the testing experience. The findings suggest that although EnglishScore holds promise as a flexible assessment tool, further refinement is necessary to enhance its reliability and acceptance as a formal academic evaluation. This research contributes to the growing discourse on digital language assessments and provides insights for educators and policymakers seeking to implement effective, technology-driven evaluation methods.
Purevdolgor Luvsantseren, Ajnai Luvsan-Ish, Javzmaa Tsend, Oyuntsetseg Sandag, Baasandorj Chilhaasuren, Odgerel Boldbaatar, Akhit Tilubai, Nyamdavaa Uugandaavaa and Darambazar Gantulga
This To reform the Bachelor of Health Informatics curriculum in Mongolia, a comparative analysis was conducted with similar international programs, and the revised curriculum was developed based on satisfaction surveys from both graduates and employers. A satisfaction survey was conducted in which 49 graduates evaluated the curriculum using a 19-question questionnaire, and 14 employers assessed the graduates’ competencies through an 8-question questionnaire, forming the basis for two distinct datasets. Factor analysis (PCA and EFA with Varimax rotation) was performed on the resulting dataset. The analysis identified five latent factors from the graduates’ evaluations—Quality of Instruction, Practical Training and Skills, Learning Environment and Resources, Curriculum Structure, and Professional Development Opportunities—and one factor from the employers’ assessments: Job Requirements and Skills Evaluation. Based on these findings, the revised curriculum was developed through a comparative analysis with programs from the University of Sydney, Oregon Tech, and Montana Tech. The updated curriculum consists of 120 credit hours and complies with Order A/147 of the Minister of Education and Science, as well as the IT Curriculum Guide 2017. A notable feature of the revised program is its emphasis on specialization in health data processing, information systems architecture, and computer networking. This research contributes to enhancing the quality of Health Informatics education and aligning the academic training with the needs of graduates and labor market demands.
Uranchimeg Tudevdagva, Ulam-Orgil Choijiljav and Erdenekhuu Norinpel
This paper describes the evaluation process of Mongolian University of Science and Technology (MUST) and its data analyze based on the structure-oriented evaluation (SURE) model. The survey questions are divided into six main groups. Each group has sub questions which measures qualitative units from very bad to very good in six levels. By this survey data is collected from all nine schools of the MUST in each semester. In this paper explains data analyze of Winter semester 2023/2024 by the descriptive statistics methods and by the SURE data processing methods. In total 14993 students are responded to the survey voluntarily. The findings showed that quality of the survey questions need to be improved based int international accreditation criteria. By the comparison of data processing two methods found out that school ranking by mean cannot be confident by the SURE model. Therefore, for the next round of the evaluation need to apply at least three different data processing methods and need to do comparison analyzes to check and proof the final results of the school’s evaluation and ranking.
Otgonchimeg Choidorj, Ariunbolor Purvee and Erdenetsetseg Saijaa
This study presents a comparative analysis of two simulation approaches—the direct-quadrature (dq) transformation and the winding function method—for modeling a squirrel cage induction motor, with the goal of minimizing the deviation between simulated outputs and the motor’s nameplate parameters. Grey Relational Analysis (GRA) was employed to optimize six key input parameters, enhancing the accuracy of the motor’s dynamic behavior representation. Simulation models were developed in MATLAB/Simulink, and key performance metrics—rotational speed (RPM), torque, and stator current—were evaluated under both steady-state and transient conditions. The optimized simulations showed a high degree of agreement with the nameplate parameters, achieving 98.5–99% accuracy. Based on these results, frequency spectrum and time-domain analyses further validated the accuracy of both simulation methods against field measurements. The winding function-based simulation proved particularly effective in identifying fault characteristic frequencies for early fault detection, while the dq-based simulation was more suitable for motor drive development. This paper also presents time- and frequency-domain analyses of the winding function-based model under various motor fault conditions, along with a comparative study of control strategies for squirrel cage and wound rotor induction motors using the dq-based model.
Sugir Tsagaanchuluun, Zagarzusem Khurelbaatar, Dorj Byambaa, Uuganchimeg Ganchimeg and Heemin Park
Self-balancing two-wheeled robots work the inverted pendulum principle, wherein the robot must move forward when it tilts forward to maintain stability, and backward when it tilts backward to avoid falling. To implement this, two sensors were used to accurately estimate and record the robot’s angular velocity. The raw outputs from these sensors were fused using a complementary filter, enabling the generation of a reliable estimate of the system’s orientation, suitable for use by control algorithms such as a proportionalintegral- derivative (PID) controller. In conjunction with this controller, a direct current (DC) motor equipped with an encoder was utilized to provide precise motion control. The pulse-width modulation (PWM) signal driving the DC motor was dynamically adjusted by the PID controller to minimize the error between the desired and actual tilt angles, thereby ensuring the robot remains upright. In this study, data generated from the PID-controlled system were collected to train both feedforward neural networks (FNN) and recurrent neural networks (RNN), enabling the development of datadriven models for system identification and control. Consequently, a NN can be trained to emulate the behavior of the PID controller, enabling it to function as a model-free, datadriven alternative for control tasks.
Narantsetseg Yadmaa, Bayar-Erdene Lkhagvasuren and Serchmaa Tserenbaljir
This study examines the concept of a digital university. It analyzes the readiness of the Mongolian University of Science and Technology (MUST) to transition from a traditional to a digital university model. Digital universities play a vital role in the sustainable development of nations by expanding access to high-quality education and fostering innovation. Mongolia, one of the world’s largest countries by land area but sparsely populated, stands to benefit significantly from digital education initiatives. MUST aims to become a national leader in digital higher education. This paper evaluates the university’s digital transformation readiness using five key dimensions, assessed through qualitative methods. The findings indicate that the strongest area is technological infrastructure, which serves as a major enabler for transformation. However, the least developed dimension is the institutional culture, skills, and human capacity required for a digital university. To successfully advance, MUST should focus on building a digital culture, enhancing digital competencies among faculty and students, and investing in human resources with digital expertise.
Dimitrios Datsogiannis and Wolfram Hardt
This paper presents the research results of using a recommendation system (RS) to evaluate automotive development processes from the initial design phases to the final software release. The model uses a dynamic questionnaire managed by a recommender system, which aims to detect deficiencies and faults during the development while increasing transparency among stakeholders. The question pool with 137 questions, consists of weighted questions, derived from key development metrics and serves as the basis for this evaluation. Questions are selected during each development cycle with Contextual Multi- Armed Bandit (CMAB) algorithms, allowing for minimal stakeholder effort while maximizing feedback relevance. The current study explores the system’s performance across four behavioral change scenarios that simulate different rates of software fault occurrence. Three reinforcement learning algorithms, eGreedy, LinUCB, and Thompson Sampling, were applied and compared. The tuning of the algorithms is also part of the performance assessment. The results are assessed based on the defined performance indicators.The findings indicate that the system is not only scalable but also achieves improved performance over time as it continues to operate and adapt to changing development conditions.
Jean Rosemond Dora, Ladislav Hluchy and Michal Stano
Traffic in network security is a topic that consistently draws attention. Companies implement several robust security measures to maintain a high standard of security within their environment. Monitoring and blocking anomalies are among the features they support. Moreover, security analysts configure network filters to minimize the probability of being exploited. We will examine the challenges that these security measures can pose to attackers. A thorough review of its utilization will also be addressed. Since these security solutions may hinder our activities, we will discuss their weaknesses and strengths, as well as their performance. In a hardened network infrastructure, i.e., that was built with security in mind, manipulating those measures is complex. Notwithstanding, it is of high interest to an attacker to defeat them. We will describe and explain the strategies that, if addressed, can circumvent these solutions.
Ghita Ikmel, Shadi Saleh, Wolfram Hardt and Najiba El Amrani El Idrissi
Object detection plays a central role in computer vision, serving as a key technology in fields such as autonomous driving, medical imaging, and robotics. This survey explores the evolution of object detection, tracing its path from traditional feature-based techniques to advanced deep learning and Transformer-based models. It provides a thorough comparison of different approaches, discusses essential evaluation metrics, and highlights critical challenges such as occlusion, scale variance, and real-time processing. Additionally, the paper reviews recent trends, including anchor-free and open-vocabulary detection, and introduces cutting-edge models like YOLOv9, RT-DETR, and Grounding DINO. The goal is to present a unified, up-to-date resource for researchers and practitioners in the field.
