Volume 14 | Issue 4

 Abstract

Because digital content is growing so quickly, sentiment analysis (SA) is now an important tool for figuring out how people feel and sorting through text data. Natural language processing (NLP) has come a long way, but low-resource languages, especially Sindhi, still haven't been studied enough because there aren't enough computational tools and annotated datasets. This study fills this gap by presenting the Sindhi News Headlines Dataset (SNHD), a new collection of data that has been labelled for both SA and category classification in eight areas: Crime, Economy, Entertainment, Health, Politics, Science & Technology, Social, and Sports. We compare different machine learning (ML), deep learning (DL), and transformer-based methods on SA and category classification tasks to see how well they work. We also use Explainable Artificial Intelligence (XAI) methods like Local Interpretable Model-Agnostic Explanations (LIME) to learn more about how models make decisions. The SNHD dataset shows that traditional ML models work better than DL and transformer-based models in experiments. Support Vector Machines with Radial Basis Function (SVM-RBF) is the best for SA (0.74 accuracy and weighted F-score), and the Ridge Classifier (RC) is the best for category classification (0.84 accuracy and weighted F-score). XLM-RoBERTa is one of the best transformer models for category classification, with an accuracy of 0.82 and a weighted F-score. These results set a standard for future research in Sindhi NLP and show how hybrid methods could help with problems that come up with low-resource languages. This work is a basic resource for NLP researchers who want to improve computational methods for Sindhi and other lesser-known languages.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Sentiment Analysis, Sindhi News Headlines Dataset (SNHD), Category-Based Classification, Machine Learning, Transformer Models, Explainable Artificial Intelligence, Low-Resource Language Processing

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

ecognising emotions is becoming more and more important for making interactions between people and computers better. This is because emotions are a big part of how people interact with each other and how they feel overall. Many industries need machines that can pick up on and respond to emotional cues like people do. Emotionally responsive agents are useful in many fields, such as education, healthcare, gaming, marketing, customer service, human-robot interaction, and entertainment. This study investigates the potential for improving virtual assistants through multimodal Artificial Intelligence (AI), employing diverse emotion recognition techniques to develop more empathetic and efficient systems. The suggested method uses facial expressions and written cues to make the system more aware of emotions and make users happy by having empathetic conversations. The Facial Emotion Recognition (FER) model was 71% accurate in real time, and the Textual Emotion Recognition (TER) model was 59% accurate in validation, showing that Multimodal Emotion Recognition (MER) works well. Our lightweight architecture makes sure that inference happens in real time and that facial and textual emotion recognition are combined with DialoGPT-based response generation. This shows that it works with large language models for empathetic dialogue, unlike previous multimodal emotion-aware systems.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Multimodal Emotion Recognition; Facial Emotion Recognition; Textual Emotion Analysis; Affective Computing; Human-Computer Interaction; Empathetic Virtual Assistants;

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Web Text Classification, BERT (Bidirectional Encoder Representations from Transformers), BiGRU (Bidirectional Gated Recurrent Unit), Convolutional Neural Network (CNN), Attention Mechanism, Natural Language Processing,

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Because higher education institutions are changing so quickly to digital, students and university administration need to be able to talk to each other quickly and easily. Students often want to know about things like how to sign up for classes, how to pay for them, deadlines, academic rules, and other administrative tasks. But traditional manual systems often slow things down, make things less consistent, and give administrative staff more work to do. This project suggests using a chatbot to help students with questions about paying for and enrolling in college (CSM). The chatbot uses Machine Learning (ML) and Natural Language Processing (NLP) to understand and answer student questions quickly and accurately. The system offers automated help around the clock, making sure that students always get accurate, timely answers without having to rely on administrative staff. There are many parts to the chatbot, such as user interaction, natural language processing, machine learning adaptation, database management, response generation, and analytics tracking. These modules work together to understand what users are asking, get the right information from institutional databases, and give answers that are relevant to the situation. To check how well the system worked, we looked at usability metrics like response time, task completion time, and user satisfaction levels. Students said they had good experiences with the chatbot because it was easy to use, fast, clear, and they felt confident using it again. The solution cuts down on repetitive work for staff by a lot, and it also makes it easier for students to access and enjoy their work. The proposed chatbot system makes administration more efficient, makes sure that information is consistent, and makes the student experience better through smart automation.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Chatbot System, Higher Education, Natural Language Processing (NLP), Machine Learning (ML),

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Contribution: This study examines the advantages and obstacles associated with the development, implementation, and assessment of a large language model (LLM) chatbot, MoodleBot, within computer science educational environments. It shows how LLMs could be used in LMSs like Moodle to help with self-regulated learning (SRL) and help-seeking behaviour. Computer science teachers have a lot of trouble adding new tools to LMSs to make the learning environment more supportive and interesting. MoodleBot solves this problem by giving students and teachers a place to interact with each other. Questions for Research: This study examines two questions, notwithstanding challenges such as bias, hallucinations, and the reluctance of teachers and educators to adopt new AI technologies. (RQ1) How much do students think MoodleBot is a useful tool for helping them learn? (RQ2) How accurate are the answers that MoodleBot gives, and how well do they fit with the course content that has already been set? Methodology: This study examines pedagogical literature regarding AI-driven chatbots and employs the retrieval-augmented generation (RAG) methodology for the design and data processing of MoodleBot. The technology acceptance model (TAM) looks at how much users accept something by looking at things like how useful they think it is and how easy it is to use. Forty-six students took part, and thirty of them filled out the TAM questionnaire. Results: Chatbots that use LLM, like MoodleBot, can make teaching and learning a lot better. This study found that there was a high accuracy rate (88%) in helping with course-related tasks. Students' positive feedback shows that AI-powered educational tools work and can be used in real life. These results show that educational chatbots can be used in courses to make learning more personalised and make teachers' jobs easier, but automated fact-checking needs to get better.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Artificial Intelligence, Large Language Models, Retrieval-Augmented Generation, Learning Management Systems, AI Chatbot, Personalized Learning, Higher Education

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

People all over the world are now paying attention to the deaths of mothers and children. In low- and middle-income countries, maternal mortality is high, especially among teens and young adults. Healthcare professionals can use CTGs to keep an eye on the mother's heartbeat during pregnancy to make sure the baby is still alive and avoid these deaths. This study utilised machine learning techniques to conduct a risk factor analysis aimed at decreasing child and maternal mortality. This study assessed seven machine learning algorithms. Accuracy, precision, and recall were used to compare how well different categorisation algorithms worked. The random forest is the most accurate of the other algorithms, with an accuracy rate of 99.98%. At first, the dataset was not balanced. After using under sampling and oversampling methods, all of the algorithms worked very well. A primary objective of the current study was to forecast the risk factors associated with child and maternal mortality utilising clinical data. Ultrasound devices work by sending out a pulse and reading the response. This analysis is a good and cost-effective choice for healthcare professionals who want to keep mothers and children from dying.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Maternal Mortality, Child Mortality Prediction, Machine Learning, Random Forest, Cardiotocography (CTG), Risk Factor Analysis, Clinical Data.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Wi-Fi-Based Human Activity Recognition (HAR), Channel State Information (CSI), Deep learning

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The modernisation of the traditional energy grid into an integrated platform is made easier by constant communication and advances in information technology. The Internet of Things (IoT) includes power systems, especially smart grid features and the ability for utilities to send new services to end users over a two-way communication channel. But relying too much on IoT-based communication systems has made security holes very serious. Also, cybercriminals are always interested in stealing important information from two people or devices, especially if they can do so by damaging the integrity, confidentiality, and authenticity of a communication channel for financial gain. Maintaining data security and preserving privacy in between two entities during the transmission or any data distribution are essential. To build a strong cyber security system, we need to look into the possible attacks and their effects. A lot of researchers have focused on finding and stopping these weak cyber attacks using advanced computing tools. This review article thoroughly investigated possible ways to address cyber security challenges such as smart meter security, end-users privacy, electricity theft cyber-attacks using blockchain and cryptography against communication attacks in smart grid. A lot of research has been done on how cyberattacks affect the security of power systems and how they affect the economy of deregulated energy markets. The resilience of security features and cryptographic techniques against diverse cyber-attacks is examined to propose uncharted cyber-attack avenues for future exploration. Specially, the study of real-world cyber security events, case studies, new findings and new scopes in diverse power industries are carried out. This review article has looked at more than 135 research papers. This paper primarily focuses on distribution-side cyberattacks, encompassing impact analysis, detection, and protection techniques.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Cybersecurity, Machine Learning, Deep Learning, Intrusion Detection Systems, Network Security

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Recognising emotions is becoming more and more important for making interactions between people and computers better. This is because emotions are a big part of how people interact with each other and how they feel overall. Many industries need machines that can pick up on and respond to emotional cues like people do. Emotionally responsive agents are useful in many fields, such as education, healthcare, gaming, marketing, customer service, human-robot interaction, and entertainment. This study investigates the potential for improving virtual assistants through multimodal Artificial Intelligence (AI), employing diverse emotion recognition techniques to develop more empathetic and efficient systems. The suggested method uses facial expressions and written cues to make the system more aware of emotions and make users happy by having empathetic conversations. The Facial Emotion Recognition (FER) model was 71% accurate in real time, and the Textual Emotion Recognition (TER) model was 59% accurate in validation, showing that Multimodal Emotion Recognition (MER) works well. Our lightweight architecture makes sure that inference happens in real time and that facial and textual emotion recognition are combined with DialoGPT-based response generation. This shows that it works with large language models for empathetic dialogue, unlike previous multimodal emotion-aware systems.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Recognising emotions is becoming more and more important for making interactions between people and computers better. This is because emotions are a big part of how people interact with each other and how they feel overall. Many industries need machines that can pick up on and respond to emotional cues like people do. Emotionally responsive agents are useful in many fields, such as education, healthcare, gaming, marketing, customer service, human-robot interaction, and entertainment. This study

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

As cyber attacks on industrial control systems become more common, it is more important than ever to use cyber security controls and check security against these attacks. Cyber attacks on nuclear power plants (NPPs) can cause not only economic loss, but also loss of life. So, to protect NPPs and other places from security threats, cyber security controls must be put in place. However, there aren't many resources available right now for protecting information, which is necessary to use all the controls needed to follow cyber security rules. To solve this problem, we need to find good cyber security controls and give each NPP enough resources to protect their information. NPPs use a different security control based on NEI 13-10 (Cyber Security Control Assessments) to protect their systems. However, this is not enough to show that the security controls have really reduced these threats or to show that they have really reduced these threats. To solve this problem, the Electric Power Research Institute (ETRI) came up with the technical assessment methodology (TAM), which can be used to give a quantitative score by looking at how possible cyber attacks could affect an asset and the security controls that go with it. This method lets you use differential security control based on the score to see if the security controls have really reduced the risks. In light of this context, the objective of this paper is to perform a comparative analysis of the outcomes obtained from the implementation of security controls and risk assessment utilising solely NEI 13-10, as well as both NEI 13-10 and TAM, on the plant protection system of the nuclear power reactor APR1400. This paper also talks about the areas for future research by talking about the TAM's limits and things to think about when using it.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Cybersecurity, Nuclear Power Plants (NPP), Technical Assessment Methodology (TAM), Risk Assessment, Industrial Control Systems (ICS).

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Healthcare professionals now frequently use medical documents due to the expansion of the Internet. To facilitate collaboration while safeguarding private patient data, secure medical image transmission and management are crucial. This study examines several approaches to securely exchanging medical data, emphasising both their benefits and drawbacks. We divide these strategies into two categories: distributed strategies like blockchain and federated learning, and centralised strategies like encryption and watermarking. This study also looks at how medical image watermarking has changed over time, from conventional approaches to sophisticated AI-based systems. Deep learning models are regarded as "black boxes," providing more resilience and flexibility than conventional methods known as "white boxes," which are straightforward and easy to understand. In order to handle the increasing complexity of threats while maintaining the diagnostic integrity of medical images, this analysis highlights the necessity of integrating contemporary technology. Additionally, our work contributes to the ongoing discussion on improving data security in medical imaging by offering a thorough classification of watermarking techniques and outlining future research directions.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Medical Image Security, Digital Watermarking, Blockchain-Based Traceability, Federated Learning, AI-Driven Tamper Detection

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

This study examines the correlation between the Latin American Artificial Intelligence Index (ILIA) and the Web Accessibility Index (WAIN) across 19 Latin American nations to assess the alignment of technological innovation with digital inclusivity. The ILIA gives a broad picture of how ready AI is by looking at things like infrastructure, data capacity, talent development, research and adoption, and governance structures. This index shows that some regions are better at digital transformation than others. For example, Uruguay, Chile, and Brazil are still ahead of the game when it comes to AI adoption because they have strong infrastructure and governance frameworks.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Artificial Intelligence Readiness (ILIA), Web Accessibility Index (WAIN), WCAG 2.2 Compliance, Digital Inclusivity, Sustainable Digital Transformation

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

As more and more autonomous and smart vehicles are used in ground transportation systems, new security problems arise. This change from operations run by people to those run by computers makes it easier for bad people to attack. As the Internet of Things (IoT) becomes more common in cars, they are always making and sharing a lot of data. Attackers can take advantage of this trend's weaknesses in complicated ways, like with Advanced Persistent Threats (APT). It is very important to be able to find APTs in vehicles that have IoT technology. To find these threats, we need better ways to do so. The urgent requirement for vehicle data privacy constrains conventional centralised Machine Learning (ML) methodologies. Also, there aren't many APT datasets available to the public in the vehicle field, which makes it harder to develop and test models. This is a big problem for cybersecurity in this area that is always changing. This study introduces an innovative Federated Deep Neural Network (FDNN) framework incorporating a privacy-preserving technique to address these challenges. The research emphasises the primary obstacles in APT detection and delineates its distinctive contributions to the domain. It talks about the research questions that are guiding the study. The UNSW-NB15, Edge-IIoTset, and CSE-CIC-IDS2018 datasets each represent a different stage of an APT attack. We use these datasets to look at and judge the framework that was made. For these datasets, the framework without the privacy-preserving technique gets APT detection accuracies of 97.32%, 96.81%, and 98.06%, respectively. But when the privacy-preserving technique is used, the framework's accuracies are 95.62%, 96.11%, and 95.63%, respectively. There are tables that show all of the results, as well as other evaluation metrics like Precision, False Positive Rate, and F1 Score. We use "Shapley Additive Explanations (SHAP)" analysis on the framework we made to find the most important features for finding APTs. This study validates the efficacy of a novel framework for identifying APTs in distributed vehicular contexts. The framework works well because it cuts down on the amount of data and the number of features. This was shown by extensive testing with several benchmark datasets. Future work will look into how well the framework can find APTs in different areas.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Federated Learning, Advanced Persistent Threats (APT), Connected Vehicles Security, Explainable Artificial Intelligence (XAI), Privacy-Preserving Deep Learning.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Tool condition monitoring (TCM) is very important in today's manufacturing because it makes sure that products are of high quality, cuts down on unplanned downtime, and makes production as efficient as possible. Over time, mechanical stress, heat, and friction cause cutting tools like milling cutters, drills, and lathes to wear down. Tool wear not only affects the size and surface quality of parts made, but it also raises production costs and can put workers' safety at risk. Most traditional TCM systems use manual inspection or supervised machine learning methods, which need a lot of labelled datasets. However, labelling data from industrial sensors takes a lot of time and work, and it is often not possible in real-world manufacturing settings, especially for machines that run on a continuous production schedule.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Tool Wear Recognition, Unsupervised Learning, Semi-Supervised Learning, Sensor Fusion, Predictive Maintenance.

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

A significant amount of web text data, including news articles, blogs, social network status updates, online reports, and more, has been gathered along with the growth of the internet. Accurately classifying these web texts automatically is a crucial task in the fields of artificial intelligence and natural language processing. Numerous applications, including sentiment analysis, web filtering, automatic recommendation, and web information retrieval, are made possible by it. The issue is that it is challenging to accurately classify web text because of its complex semantic structures and distant relationships. Current word representations, like Word2Vec and GloVe, generate fixed word embeddings for each and every word. This paper presents a more advanced deep learning structure called BERT, BGCA, which combines BERT (Bidirectional Encoder Representations from Transformers), BiGRU (Bidirectional Gated Recurrent Unit), CNN (Convolutional Neural Network), and Attention in a web text classification problem to address the aforementioned problems. In order to help the model learn the syntax and semantics between words through bidirectional transformer architecture, BERT is used to obtain the contextual encoding of words in the web text. Contexts can be investigated both forward and backward using the BiGRU layer. The CNN layer is in charge of extracting important details and gram features from the text. and the focus is on capturing the importation of key terms. To confirm the effectiveness of the suggested model, BERT, BGCA, experiments were carried out on a sizable enough news dataset called THUCNews. Conventional embedding techniques (Word2Vec and GloVe) and BERT-based embedding were tested. According to the comparison results, the BERT significantly outperformed the static embedding techniques. On the 20 categories THUCNews news dataset, the high accuracy and F1 score (95.21%, 94.36%) and F, measure of 95.21% and 94.36%, respectively, demonstrate that the BERT, BGCA can achieve better classification results and be favourable to semantic representation when compared to other deep learning models like TextCNN. Therefore, it can be said that the BERT, BGCA method for large-scale web text classification is practical, scalable, and effective.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Web Text Classification, BERT (Bidirectional Encoder Representations from Transformers), BiGRU (Bidirectional Gated Recurrent Unit), Convolutional Neural Network (CNN), Attention Mechanism, Natural Language Processing,

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Working capital management is a very important part of financial management, which involves the management of short-term assets and liabilities of an organisation. This includes maintaining sufficient liquidity to conduct organisational activities in an efficient manner. In addition to this, this involves improving the profitability of the organisation. Working capital management plays an important role, especially in the manufacturing sector, because in such industries most of the funds are invested in inventories and receivables. In this study, working capital management strategies followed by JALARAM DOORS & HARDWARES, a manufacturing and trading organization involved in wooden doors manufacturing and hardware supply, have been studied. The study relies upon secondary data collected from financial statements of the organisation from 2021-22, 2022-23, and 2024-25. Financial ratio analysis will be conducted through the current ratio, inventory turnover ratio, and debtor's turnover ratio. From the findings, it is seen that although there are satisfactory levels of liquidity in the organization, there are some issues associated with the management of inventories and fluctuation in the profitability levels

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Working Capital, Liquidity, Inventory, Receivables, Profitability

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

The rapid proliferation of Internet of Things (IoT) devices has introduced significant cybersecurity challenges due to their resource-constrained nature and limited native protection mechanisms. Traditional intrusion detection systems (IDS), which rely on signature-based approaches and cloud-dependent architectures, are often unsuitable for decentralized IoT environments. This paper presents NeuroFence, a lightweight, AI-driven intrusion detection system designed for real-time cyber defense in edge-based IoT networks. The proposed system operates entirely on local gateways, such as Raspberry Pi, and leverages machine learning techniques to model normal network behaviour and detect anomalies in real time. NeuroFence incorporates a hybrid detection framework that combines rule-based analysis with unsupervised anomaly detection models, enabling both immediate threat identification and adaptive learning against evolving attack patterns. The system captures live network traffic, extracts features using sliding-window techniques, and evaluates traffic behaviour using models such as Isolation Forest and TinyML-based classifiers. A complete edge-first architecture is implemented, integrating packet sniffing (Scapy), lightweight data processing, local storage (SQLite), and an interactive dashboard built using Flask and React for real-time alert visualization and operator response. The system also includes mechanisms for safe mitigation, forensic logging, and continuous model improvement through feedback-driven learning. Experimental evaluation demonstrates that NeuroFence effectively detects anomalous IoT traffic with low latency and minimal computational overhead, making it suitable for deployment in resource-constrained environments such as smart homes, industrial IoT systems, and remote installations. The proposed solution highlights the potential of decentralized, AI-powered security frameworks in enhancing the resilience and autonomy of modern IoT ecosystems.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Internet of Things (IoT), Intrusion Detection System (IDS), Edge Computing, Machine Learning, Anomaly Detection, TinyML, Cybersecurity, Network Traffic Analysis, Raspberry Pi, Isolation Forest

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Land use pattern is a reflection of the interrelationship between natural resources and human activities in a region. Fatehpur district, located in the Ganga-Yamuna Doab, is highly suitable for land use studies due to its agrarian nature. The present study is based on the spatio-temporal analysis of land resources in Fatehpur district. The main objective of the study is to understand the distribution, trends, and variability of different land use categories in the district--such as agricultural land, forest area, fallow and wasteland, pasture land, and built-up area. In this research, secondary data (district statistical handbook, land records, and census data) have been used to classify land use and analyze changes over time. The analysis clearly indicates that agriculture is the dominant land use in the district, occupying the largest proportion of the total area, whereas forest area is very limited. Due to urbanization and infrastructural development, there has been a gradual increase in built-up areas, which is putting pressure on agricultural land. The study also reveals that land use patterns are strongly influenced not only by physical factors--such as soil, climate, and water resources--but also by socio-economic factors like population growth, technological development, and market accessibility. Additionally, issues such as land degradation, declining groundwater levels, and imbalanced land use are emerging challenges in the region. Finally, the study emphasizes the need for sustainable land use management measures, such as agricultural diversification and the promotion of modern irrigation techniques, to ensure balanced and long-term utilization of resources in Fatehpur district.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Land Use Pattern, Spatio-Temporal Analysis, Fatehpur District

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Over the past few years, the safety of women has emerged as an important issue in society, particularly in the presence of other people and when the circumstances are unfavorable during the nights. As the number of harassment and emergency incidents increases, the need to find intelligent, non-obtrusive safety solutions that will work in real time is growing. The proposed project is called Night Guardian: Empowering Women Safety with Hand Sign-based Communication and AI and suggests a new solution that would be based on leveraging artificial intelligence and computer vision to provide immediate assistance in case of distress scenarios without having to engage with the user directly. The system is developed based on the latest technologies like MediaPipe and OpenCV to record and process real-time video feeds of a webcam. MediaPipe is applied in the correct hand landmark detection and keypoint extraction in order to determine hand gestures very precisely. These features are extracted and converted into numerical forms and input to machine learning algorithms like Random Forest, Support Vector Machine (SVM) and K-Nearest Neighbor (KNN). Performance measures such as precision, recall, and F1-score are used to train and test the models to be reliable at gesture recognition. After a distressing gesture has been identified, an alert system is automatically triggered. This involves raising a local alarm to draw attention locally and an automated email notification and a picture of the gesture to emergency contacts or authorities that are predetermined. The combination of these alert mechanisms is to make sure that there is quick response hence increasing the possibility of intervention in time. With the application of AI, the system can be continuously and effectively used even in dynamic environments. Altogether, the offered system will be an effective and scalable way to enhance the safety of women with the help of intelligent automation. It offers a quick, convenient, and free form of emergency communication by using machine learning and real-time gesture recognition to deliver a hands-free form of communication. This project, apart from fostering innovation in technology, helps to enhance social well being by making people be empowered with proactive safety tools.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Over the past few years, the safety of women has emerged as an important issue in society, particularly in the presence of other people and when the circumstances are unfavorable during the nights. As the number of harassment and emergency incidents increases, the need to find intelligent, non-obtrusive safety solutions that will work in real time is growing. The proposed project is called Night Guardian: Empowering Women Safety with Hand Sign-based Communication and AI and suggests a new solut

  License

Creative Commons Attribution 4.0 and The Open Definition

 Abstract

Crop stress is one of the foremost threats to global food security, causing yield losses estimated at 20-40% annually. Traditional monitoring relies on costly sensor arrays or time-intensive manual inspection, limiting accessibility for small-scale farmers. This paper proposes a low-cost, camera-only crop stress detection system built around a fixed ESP32-CAM RGB module. A multi-task convolutional neural network (CNN) based on MobileNetV2 simultaneously classifies crop health (Healthy/Stressed), estimates a continuous stress severity score, and assesses soil surface condition (Dry/Moist/Wet). Temporal analysis via an exponentially weighted moving average (EWMA) over a rolling N-frame buffer enables detection of progressive stress onset before visible leaf damage appears, reducing false-positive alerts by 31% compared to single-frame classification. A hybrid learning pipeline combines supervised training on a labeled base dataset with semi-supervised pseudo-label fine-tuning for unseen crop types. Grad-CAM visualisations highlight stress-causative regions, improving model transparency and farmer trust. The system achieved 93.4% binary classification accuracy, 89.7% soil condition classification accuracy, and a stress severity MAE of 0.06 on held-out test data.

Licence: creative commons attribution 4.0

  License

Creative Commons Attribution 4.0 and The Open Definition

 Keywords

Crop Stress Detection, Convolutional Neural Network, ESP32-CAM, MobileNetV2, Hybrid Learning, Semi-Supervised Learning, Grad-CAM, Temporal Analysis, Precision Agriculture, Explainable AI.

  License

Creative Commons Attribution 4.0 and The Open Definition