Agentic Vision Framework for Real-Time Manufacturing Contamination Detection Using Patch-Based Lightweight Convolutional Neural Networks

This paper presents a novel framework for detecting and predicting abnormal traffic events on highways. Traditional traffic monitoring systems often rely on a single data source, which limits detection accuracy and robustness in complex environments. To address these challenges, we propose a multimodal deep fusion framework based on heterogeneous graph neural networks (HGNNs), enhanced by an ensemble contrastive pessimistic likelihood estimation (CPLE) algorithm. The framework integrates both static and dynamic traffic data, including video images, traffic flow, vehicle speed, and tunnel weather conditions. Through effective feature fusion, it significantly improves the accuracy and real-time performance of anomaly detection. Experimental results show that the model performs robustly across various scenarios, accurately identifying abnormal traffic events with high precision and stability. Compared with existing models such as AGC-LSTM and AttentionDeepST, the proposed MHGNN-CPLE model demonstrates superior performance, particularly in static detection tasks, achieving an accuracy of 0.980 and an F1 score of 0.967. In contrast, AGC-LSTM and AttentionDeepST achieve 0.965/0.945 and 0.960/0.935 in accuracy and F1 score, respectively. In dynamic scenarios, the model also maintains high accuracy under varying noise levels, indicating strong robustness. The research is motivated by the growing challenges of urbanization, where real-time detection and prediction of traffic anomalies are increasingly critical. Our framework effectively integrates multimodal data and leverages HGNNs to capture complex spatiotemporal dependencies, while the CPLE algorithm enhances robustness under uncertainty. The results confirm that the proposed method offers a reliable and accurate solution for real-time traffic anomaly detection, representing a significant advancement in intelligent transportation systems.

Highly motivated, sophisticated cyberattacks that target cloud capabilities with Internet of Things integrations require advanced, real-time, intelligent, scalable, and efficient Intrusion Detection Systems. This research proposes a novel Intrusion Detection Network (IDNet) architecture that utilises deep learning frameworks, combining Bidirectional Gated Recurrent Units (Bi-GRUs) with Attention Mechanisms to construct complex temporal dependencies in traffic while emphasising critical traffic instructions, pattern recognition, and other vital tasks. IDNet was implemented and tested on the Coburg Intrusion Detection Data Sets (CIDDS), which demonstrated superior performance compared to GRU, Attention-GRU, and Bi-GRU baselines in terms of accuracy and robustness. The proposed pipeline is implemented using Kubeflow Pipelines for Model training automation, Katib for hyperparameter optimisation, and MLflow/Kubeflow Metadata for model version control. Real-time inference is served using IDNet’s deployment on KServe, and performance is optimised with TorchServe and TensorRT. Through Grafana and Prometheus, observability is continuous and dynamic for metrics such as latency, throughput, false positive rate, marker shedding, and others. Adaptation to new changes is facilitated by the Population Stability Index, which initiates automatic retraining, ensuring defence against emerging threats. The solution is built on and works with Amazon Elastic Kubernetes Service. Integrated with Kafka/NATS, real-time traffic ingestion uses them for injection. After retraining the FPR IDS, IDNet achieved 98.90% accuracy alongside a 43.68% drop in latency and a 37.19% reduction in FPR. The data shows experimental results. Using comparative evaluation with the most advanced existing models validates IDNet’s efficiency as a real-time Intrusion Detection System for complex and high-traffic network environments.

Real-time object detection on resource-constrained edge devices remains a critical challenge in precision agriculture and autonomous systems, particularly when integrating advanced multi-modal sensors (RGB-D, thermal, hyperspectral). This paper introduces FEGW-YOLO, a lightweight detection framework explicitly designed to bridge the efficiency-accuracy gap for fine-grained visual perception on edge hardware while maintaining compatibility with multiple sensor modalities. The core innovation is a Feature Complexity Descriptor (FCD) metric that enables adaptive, layer-wise compression based on the information-bearing capacity of network features. This compression-guided approach is coupled with (1) Feature Engineering-driven Ghost Convolution (FEG-Conv) for parameter reduction, (2) Efficient Multi-Scale Attention (EMA) for compensating compression-induced information loss, and (3) Wise-IoU loss for improved localization in dense, occluded scenes. The framework follows a principled "Compress, Compensate, and Refine" philosophy that treats compression and compensation as co-designed objectives rather than isolated knobs. Extensive experiments on a custom strawberry dataset (11,752 annotated instances) and cross-crop validation on apples, tomatoes, and grapes demonstrate that FEGW-YOLO achieves 95.1% mAP@0.5 while reducing model parameters by 54.7% and computational cost (GFLOPs) by 53.5% compared to a strong YOLO-Agri baseline. Real-time inference on NVIDIA Jetson Xavier achieves 38 FPS at 12.3 W, enabling 40+ hours of continuous operation on typical agricultural robotic platforms. Multi-modal fusion experiments with RGB-D sensors demonstrate that the lightweight architecture leaves sufficient computational headroom for parallel processing of depth and visual data, a capability essential for practical advanced sensing systems. Field deployment in commercial strawberry greenhouses validates an 87.3% harvesting success rate with a 2.1% fruit damage rate, demonstrating feasibility for autonomous systems. The proposed framework advances the state-of-the-art in efficient agricultural sensing by introducing a principled metric-guided compression strategy, comprehensive multi-modal sensor integration, and empirical validation across diverse crop types and real-world deployment scenarios. This work bridges the gap between laboratory research and practical edge deployment of advanced sensing systems, with direct relevance to autonomous harvesting, precision monitoring, and other resource-constrained agricultural applications.

Osteoporosis is a progressive skeletal disorder characterized by decreased Bone Mineral Density (BMD) and structural deterioration, leading to an increased risk of fractures. Traditional diagnostic methods, such as Dual-Energy X-ray Absorptiometry (DXA) and the Fracture Risk Assessment Tool (FRAX), often fail to detect osteoporosis at an early stage due to their limited sensitivity and accessibility. Recent advancements in Artificial Intelligence (AI), Deep Learning (DL), and Machine Learning (ML) have shown significant potential in improving osteoporosis detection. However, existing models face challenges related to data privacy, explainability, computational complexity, and generalizability across diverse populations. This study introduces Federated Graph-based Multimodal Osteoporosis Diagnosis (FGMOD), a novel AI-driven framework that integrates Federated Learning (FL) and Graph Neural Networks (GNNs) for enhanced osteoporosis risk assessment. Unlike traditional DL models that rely on centralized training, FL ensures privacy-preserving learning across multiple hospitals and institutions. Simultaneously, GNNs model trabecular bone connectivity, offering deeper insights into microarchitectural changes in osteoporotic bones. Multimodal data fusion, incorporating structured clinical data, DXA scans, and biochemical markers, significantly improves prediction accuracy and interpretability. The proposed FGMOD model was validated on a real-world dataset comprising 10 000 patient records, 5000 DXA scans, and biochemical profiles. Comparative evaluation with state-of-the-art AI models demonstrated superior performance, achieving 94.2% accuracy, 93.4% precision, 93.1% recall, and a ROC-AUC score of 0.96. Unlike CNN-based models, which rely heavily on high-resolution images, FGMOD effectively extracts clinically relevant features from graph-based bone connectivity analysis, ensuring robust and interpretable predictions. Furthermore, the use of SHAP-based feature importance analysis provides clinicians with an explainable AI decision-making process, enhancing trust in the model’s recommendations. Performance comparisons against CNN, Gradient Boosting, Random Forest, and Transfer Learning models revealed that FGMOD outperformed all baselines, particularly in privacy-aware federated training, osteoporosis risk stratification, and real-time inference on edge AI devices. Additionally, FGMOD reduced inference latency to 45 ms per sample, making it suitable for real-time clinical deployment. FGMOD represents a significant advancement in AI-powered osteoporosis detection, addressing key challenges such as privacy, generalizability, and interpretability. Future research should focus on expanding FL collaborations across global hospitals, integrating wearable sensor data for continuous bone health monitoring, and enhancing GNN-based osteoporosis modeling for faster inference on mobile healthcare platforms. This study paves the way for clinically reliable, privacy-preserving, and scalable AI solutions in osteoporosis risk assessment and early detection.

The railway system is a green mode of transport that is essential for contributing significantly to economic growth, enhancing accessibility, and easing regional integration. The safety and efficiency of railway transport systems depend on the condition of wheels, as deterioration of wheels is a major cause of both human life and financial loss. Therefore, real-time monitoring is essential for early detection and preventing failures. This study presents an AI-based framework for real-time railway wheel defect detection, leveraging advanced You Only Look Once (YOLO) models (v5-v12) and a Real-Time Detection (RTD) Transformer model. A custom wayside imaging system was developed, capturing high-resolution images to construct the FaultSeg dataset, addressing class imbalances and annotation challenges. Eight YOLO models and the RTD Transformer were evaluated, with extensive hyperparameter tuning, to identify defects such as wheel flats, shelling, discoloration, and cracks/scratches. The YOLOv5-seg model demonstrated superior performance with 91% precision, 90% recall, and 92% mAP@0.5, achieving real-time processing at 30 FPS with latency under 30 ms. The optimized model was deployed on an edge device for operational railway environments, showcasing its feasibility for real-time defect detection, enhancing predictive maintenance, and improving railway safety. This work contributes to the advancement of AI in condition monitoring by providing a publicly available dataset and demonstrating the practical deployment of a real-time defect detection system.

While mobile social networks (MSNs) enrich people's lives, they also bring many security issues. Many attackers spread malicious URLs through MSNs, which causes serious threats to users' privacy and security. In order to provide users with a secure social environment, many researchers make great efforts. The majority of existing work is aimed at deploying a detection system on the server and classifying messages or users in MSNs through graph-based algorithms, machine learning or other methods. However, as a kind of instant messaging service, MSNs continually generate a large amount of user data. Without affecting the user experience, with existing detection mechanisms it is difficult to implement real-time detection in practical applications. In order to realize real-time message detection in MSNs, we can build more powerful server clusters or improve the utilization rate of computing resources. Assuming that computing resources of servers are limited, we use edge computing to improve the utilization rate of computing resources. In this article, we propose a multistage and elastic detection framework based on deep learning, which sets up a detection system at the mobile terminal and the server, respectively. Messages are first detected on the mobile terminal, and then the detection results are forwarded to the server along with the messages. We also design a detection queue, according to which the server can detect messages elastically when computing resources are limited, and more computing resources can be used for detecting more suspicious messages. We evaluate our detection framework on a Sina Weibo dataset. The results of the experiment show that our detection framework can improve the utilization rate of computing resources and can realize real-time detection with a high detection rate at a low false positive rate.

Money laundering hides illegal money’s origin by making it seem legal. Detecting suspicious activity quickly in financial data is key to stopping fraud and money laundering. Real-time detection is popular approach for its speed and efficiently detecting illegal activities in financial institutes system. However, handling massive and distributed data streams have challenges in achieving real-time efficiency and effectiveness. Therefore proposing and developing data stream framework is needed to handle these challenges efficiently. The main goal of this study is to propose real-time suspicious detection framework for financial institutions to effectively combat money laundering. The proposed model comprises two approaches: a distributed computing architecture based on Docker container to enhance flexibility, migration capabilities, and customization, and a suspicious detection module employing the autoencoder method. To determine whether there is any suspicious activity in the system, the proposed model uses the reconstruction error. The reconstruction error is the difference between the original input data and the data reconstructed by the proposed model. To evaluate the proposed model, we used real-world data from a financial institution and synthetic data generated from the real-world data. The study demonstrates the better performance of the proposed real-time detection framework compared to traditional methods in identifying anomalous transactions. It also explores the importance and limitations of using both real-world and generated data. Our code is publicly available: https://github.com/Ermiyas21/Real-Time-Suspicious-Detection-Framework-for-financial-data.

Detecting bacteria contamination on medical device surfaces using an integrated fiber-optic mid-infrared spectroscopy sensing method

Network-on-chip (NoC) is widely employed by multicore system-on-chip (SoC) architectures to cater to their communication requirements. Increasing NoC complexity coupled with its widespread usage has made it a focal point of potential security attacks. Distributed denial-of-service (DDoS) is one such attack that is caused by malicious intellectual property (IP) cores flooding the network with unnecessary packets causing significant performance degradation through NoC congestion. In this article, we propose an efficient framework for real-time detection and localization of DDoS attacks. This article makes three important contributions. We propose a real-time and lightweight DDoS attack detection technique for NoC-based SoCs by monitoring packets to detect any violations. Once a potential attack has been flagged, our approach is also capable of localizing the malicious IPs using the latency data in the NoC routers. The applications are statically profiled during design time to determine communication patterns. These patterns are then used for real-time detection and localization of DDoS attacks. We have evaluated the effectiveness of our approach against different NoC topologies and architecture models using both real benchmarks and synthetic traffic patterns. Our experimental results demonstrate that our proposed approach is capable of real-time detection and localization of DDoS attacks originating from multiple malicious IPs in NoC-based SoCs.

Food safety involves preparation and storage of food in different ways to prevent food-borne illness. Ever increasing incidences of food-borne diseases have led food industry to employ rapid and inexpensive method of analysis. It is important not only for health, but also from economic point of view because food-borne diseases result in financial losses. Stringent laws have been made for preparation, packaging, and storage of food. Various techniques are used in food industry but biosensing has received considerable attention due to its high specificity and quick response at low cost. The present review describes use of biosensors for the detection of various pathogens, allergens, pesticide residues, natural contaminants, and other toxic substances in food. Different types of biosensors are used to serve the purpose of food quality assurance such as electrochemical, optical, piezoelectric, and thermometric. Various materials such as monoclonal antibodies, aptamers, RNA, and DNA have been used to enhance the sensitivity and specificity of biosensors for food. Employment of nanoparticles in fabrication of biosensors has shown remarkable potential because of their unique properties at small scale. A wide variety of nanomaterials such as carbon nanotubes, nanoparticles, nanowires, and quantum dots have been used for fabrication of biosensors. Attempts have been made to achieve real-time detection of food contaminants and a little success has been obtained in this regard because more or less, almost every technique requires sample preparation. Highly sensitive and selective biosensors have been developed which showed great potential to mark the presence of food contaminants close to real-time detection. Sensing strategies are moving towards the perfection to obtain real-time detection while assuring high quality of food which is free from any type of contamination.

A real-time detection and monitoring (RTDM) of microbial contamination on solid surfaces is mandatory in a range of security, safety and bio-medical applications where surfaces are exposed to accidental, natural or intentional microbial contamination. This work presents a new device, the BC-Sense, which allows a rapid and user-friendly RTDM of microbial contamination on various surfaces while assessing the decontamination kinetics and degree of cleanliness. The BC-Sense LIDAR (Light Detection and Ranging) device uses the Laser-Induced Fluorescence (LIF) method based on dual wavelength sensing with multispectral pattern recognition system to rapidly detect microbial contamination on a solid surface. Microbial simulants (bacteria, bacterial spores, fungal conidia and virus) were spread at varying concentrations on a panel of solid surfaces which were assessed by BC-Sense. The spectra of dead and living E. coli showed differences at various sensing wavelengths. The limit of detection (LoD) of E. coli and MS2 virus was 2.9 × 104 and 9.5 × 104 PFU and CFU/cm2, respectively. Random samples (n=200) tested against a training dataset (n=800) were optimally discriminated for contamination versus background with a threshold of predicted response (PR) >0.55 and 10 min with spores and E. coli.

Mercury contamination is a global environmental issue due to its toxicity and persistence in ecosystems. It poses a particular risk in aquatic systems, where it bioaccumulates and biomagnifies, leading to serious health impacts on humans. Therefore, effective detection technologies for mercuric ions in natural water resources are highly desirable. However, most existing detection methods are time-consuming, require complicated sample pre-treatment, and rely on expensive equipment, which hinders their widespread use in real-time detection. Here, we present a convenient, rapid, portable, user-friendly, and cost-effective sensing system for detecting Hg2+ ion contamination in water. This system utilizes a highly selective, amphiphilic, and structurally simple molecular probe, N-dodecylamine-di-thiocarbamate (DDC). DDC molecules align at the interface between the liquid crystal (LC) and water, inducing a homeotropic LC orientation. In water samples contaminated with Hg2+, a bright optical texture is observed, indicating the alignment of the 5CB LC in a planar manner at the LC/aqueous boundary. The minimum detectable concentration (LOD) for Hg2+ ions is 5.0 μM in distilled water, with a broad detection range from 5.0 μM to 2 mM. The sensor selectively detects Hg2+ ions over other common interfering metal ions, including Pb2+, Co2+, Ni2+, Cu2+, Cd2+, Zn2+, Cr2+, Mg2+, Na+, K+, and Ca2+. Boolean logic gates, bar graphs, and truth tables are employed to explain the selectivity of this liquid crystal-based sensor. This work demonstrates the significant potential of the sensor for monitoring mercuric ions in natural water resources, offering a promising strategy for controlling mercury pollution.

In this pandemic-prone era, health is of utmost concern for everyone and hence eating good quality fruits is very much essential for sound health. Unfortunately, nowadays it is quite very difficult to obtain naturally ripened fruits, due to existence of chemically ripened fruits being ripened using hazardous chemicals such as calcium carbide. However, most of the state-of-the art techniques are primarily focusing on identification of chemically ripened fruits with the help of computer vision-based approaches, which are less effective towards quantification of chemical contaminations present in the sample fruits. To solve these issues, a new framework for chemical ripening and contamination detection is presented, which employs both visual and IR spectrometric signatures in two different stages. The experiments conducted on both the GUI tool as well as hardware-based setups, clearly demonstrate the efficiency of the proposed framework in terms of detection confidence levels followed by the percentage of presence of chemicals in the sample fruit.

Water pollution is a detrimental issue that occurs when there are bad changes in water quality parameters. It directly disrupts water usage and poses a danger to society, environment, economy, and agriculture. Water quality should be monitored to alert authorities on water pollution, so that action can be taken quickly. Improper water management pollutes rivers and lakes in Malaysia such as Klang River, Semenyih River, Kim Kim River, Slim River, and others. Various techniques are introduced to detect contaminants in water such as electronic sensors, biosensor approaches, laboratory analysis, and optical techniques. The functionality tool varies depending on the specific contaminants or parameters that are targeted, and the resources allocated. Several common contaminant types are found in polluted water including heavy metals, organic and inorganic chemicals, industrial pollutants, suspended solids, and others. The objective of the review is to study various non-optical and optical contaminant detection methods in water to identify the strengths and weaknesses of the methods. In this review, water pollution problems mainly due to the agricultural sector in several countries are discussed. Besides, conventional, and modern methods are compared in terms of parameters, complexity, and reliability. We believe that conventional methods are costly and complex, whereas modern methods are also expensive but simpler with real-time detection. Recent contaminant detection methods in water are also reviewed to study any loopholes in the latest methods. We found that the spectroscopy method based on light propagation theory is suitable and one of the promising methods for chemical contaminants detection for water quality analysis. This method offers fast analysis time, high sensitivity detection, non-invasive analysis, provides reliable data, and others. Undoubtedly, some contaminants in water can be challenging to be identified rapidly and confidently even though there are numerous tools and techniques available for water quality analysis. Thus, the review is important to compare previous methods and to improve current chemical contaminants detection analysis in terms of reliability with a minimum operating system and cost effectiveness.

Real-time accurate detection of wind turbine downtime - An Irish perspective