This paper proposes a machine learning model to predict the accuracy of extreme rainfall events by exploiting the concept of quality control chart in operations management. In this framework, we introduce the 3σ framework, a novel approach to short-term rainfall forecasting by presenting the rainfall process in a statistical quality control perspective. The framework consists of two parts: (1) a 3σ chart and (2) a machine learning classification model. Rainfall intensity is categorized into three classes based on the 3σ chart. The model is able to effectively capture sequential rainfall trends and predict precipitation classes up to 24 hours in advance. The results indicate that the framework achieves high performance in key metrics, including loss, precision, and recall, with consistent alignment between the training and validation phases. Furthermore, the comparison between predicted and actual rainfall classes confirms the model’s effectiveness in detecting both the occurrence and magnitude of severe rainfall events, although slight overestimations were observed in isolated cases. In general, the framework has significant potential for integration into real-time early warning systems, helping to reduce the impact of climate-driven extreme weather events by allowing faster and more interpretable alerts for floods, landslides, and related hazards.
Vision Transformers (ViTs) have achieved great performance on a variety of computer vision tasks, including medical image classification. Fine-tuning these models is computationally costly. However, Parameter-Efficient Fine-Tuning (PEFT) methods address this limitation by fine-tuning a minimal number of parameters. Despite their efficiency, the success of these methods depends greatly on the hyperparameter selection. In this paper, we propose an automatic hyperparameter search method for BEFT using Optuna.To validate the effectiveness and generality of our approach, we applied this Auto-Optimized PEFT strategy to three state-of-the-art pretrained Vision Transformers: ViT-Base, ViT-Medium, and Swin Transformer. After optimization using our method, ViT-Base achieved an accuracy improvement of +10.58% (87.17%), ViT-Medium achieved an accuracy improvement of +7.09% (83.23%), and Swin Transformer achieved an accuracy improvement of +18.62% (85.57%). The results highlight the effectiveness of PEFT methods in combination with automatic hyperparameter tuning to successfully fine-tune large vision models even in resource-scarce medical environments.
Digitalization is an initial step and foundation to any modern automation system, and a prerequisite for control of complex systems. This paper focuses on the application of machine learning methods for sensorless identification of window openness and, furthermore, window openness prediction. Identification uses data-based models developed using historical data obtained from 70 sensors installed on a single floor of the faculty building, correlated with outside meteorological conditions. Models are trained by using two machine learning approaches, Random Forest and Gradient Boosting, and the results are thoroughly analyzed and compared to each other and to experimentally obtained data. The paper also explores ways in which model training can be improved and model accuracy enhanced. The results confirm the potential of sensorless digitalization by employing machine learning algorithms with the goal of improving climate comfort and energy efficiency through automated and intelligent decision-making systems and highlight the importance of high-quality data and thoughtful algorithm selection.
This paper aims to tackle the cyber threats posed by decentralized settings, such as edge computing, by exploiting the Transfer Learning (TL) approach to create a lightweight Intrusion Detection System (IDS) for resource-constrained IoT devices. The proposed approach shifts the training complexity to the edge, allowing resource-constrained IoT devices to use less resources because they receive from the edge a single (unified) pre-trained supermodel. This supermodel ensures real-time adaptability across multiple datasets and edge environments and achieves 99% accuracy, precision, recall, and F1-score, while reducing latency and increasing scalability. Beyond TL, the proposed approach combines BERT-based semantic feature extraction, MLP classification, and SMOTE class imbalance compensation for improved performance.
Coronary artery calcification (CAC) is a strong predictor of cardiovascular events, traditionally assessed via manual scoring on ECG-gated CT scans. However, manual methods are time-consuming and subject to interobserver variability, especially in non-dedicated CT scans. This study presents a preliminary investigation into automated CAC segmentation using a dual-input W-Net architecture, which processes both raw CT volumes and thresholded binary masks (threshold: 130 Hounsfield units). The model was trained and evaluated on 93 patients, with CAC severity categorized by Agatston score. We tested both a 2-class (left/right arteries) and a 4-class (LAD, RCA, LCX, LMCA) segmentation approach. Morphological post-processing was used to refine predictions. The model’s performance was evaluated using the absolute error in Agatston score. In the 2-class model, post-processing reduced the error for the left artery from 302.7 to 51.4 and for the right artery from 1001.7 to 717.3 in the severe CAC group (>500 AS). In the 4-class model, post-processing reduced error for LAD from 99.1 to 25.7 and for RCA from 643.9 to 161.4 in the same group. Across models, mild CAC cases (<100 AS) had near-zero errors, while most false positives arose from mitral and aortic valve calcifications. These results show the potential of our approach for opportunistic CAC screening using routine CT scans. The proposed method significantly reduces absolute Agatston score error and manual burden, supporting its applicability in clinical workflows. Future work will focus on dataset expansion, model robustness, and artery-specific refinement.
The rapid proliferation of edge IoT systems in critical infrastructures, from smart cities to industrial IoT (IIoT) environments, has introduced significant security challenges, particularly Distributed Denial-of-Service (DDoS) attacks. These attacks can degrade service quality and compromise the availability and integrity of services. Although deep learning (DL) models have shown promise in detecting DDoS attacks, their reliance on large, high-quality labeled datasets limits their adaptability in dynamic IoT environments. Transfer learning offers a potential solution; however, existing methods often struggle with domain adaptation and effective knowledge transfer across heterogeneous datasets, leading to suboptimal performance against evolving attack patterns. To address these challenges, we propose TransferEdge, a novel transfer learning-based approach to detect evolving DDoS attacks in industrial IoT edge systems. TransferEdge leverages pre-trained models and describes a novel approach to optimize fine-tuning strategies tailored for DDoS attack detection, so as to align feature spaces and bridge the distributional gap between source and target domains. Experimental evaluations on the UNSW-NB15 and BoT-IoT datasets demonstrate that TransferEdge improves detection accuracy and decreases training time compared to conventional DL methods and current transfer learning approaches.
Energy consumption forecasting remains a critical challenge in power grid management, with existing methods struggling to balance computational complexity with prediction accuracy. Here, we present a novel hybrid approach that decomposes power consumption signals into frequency components before applying deep learning, addressing the limitations of current approaches in capturing both long-term trends and rapid fluctuations. By separating low-frequency trends from high-frequency components using moving average filtration, and processing each through specialized Long Short-Term Memory (LSTM) networks, our method achieves superior prediction accuracy compared to traditional single-stream approaches. The frequency decomposition plays a crucial role in improving predictive performance by allowing the model to focus separately on persistent consumption patterns and transient variations, reducing noise interference, and enhancing feature extraction. Testing on two distinct datasets – Smart House and Mexican Household – demonstrates the robustness of the methodology, yielding Root Mean Square Error (RMSE) values of 0.0132 and 0.0195, respectively. Our findings establish frequency decomposition as an effective strategy for energy forecasting, enabling better capture of both short- and long-term dependencies, with immediate applications in smart grid management and energy optimization systems.
This study addresses the integration of ergonomic risk management into production cost optimization, emphasizing the role of operators' health in sustaining efficiency and minimizing costs. Work-related musculoskeletal disorders (WMSDs), often accentuated by ergonomic stressors in repetitive and high-risk tasks, significantly impact workforce productivity and incur high direct and indirect costs for organizations. This research proposes a dual-evaluation framework, combining established ergonomic assessment tools (REBA and OCRA) with a personalized assessment of operator health characteristics via the Cox proportional hazards model. By assigning a unique ergonomic score to each operator and normalizing workstation workload based on severity, the model facilitates an optimized rotation strategy that balances ergonomic loads across operators, reducing injury risks and supporting sustained production levels. A structured production line analysis demonstrates this framework’s effectiveness in reducing ergonomic imbalances and operational costs. The findings emphasize ergonomic risk management as essential for sustainable production planning and lay groundwork for future research on cumulative ergonomic scoring for long-term health impacts.
The proliferation of Internet of Things (IoT) devices has significantly intensified cybersecurity concerns, highlighting the need for robust, adaptable, and privacy-preserving Network Intrusion Detection Systems (NIDS). A major challenge lies in the heterogeneity of IoT environments, which complicates the generalization of detection models across different network contexts. In this work, we propose a network-agnostic NIDS enhanced through Class Incremental Learning (CIL), allowing the integration of legitimate traffic from different networks without requiring retraining from scratch or exposing sensitive data. Our approach ensures efficient, continuous adaptation to new environments while maintaining strong detection capabilities. To assess the effectiveness of the proposed solution, we evaluate several CIL techniques in two deployment scenarios: within the same network and across different networks. Results show that the best-performing CIL methods perform comparably to an upper-bound model trained from scratch, with minimal knowledge degradation when adapting to previously unseen benign traffic. These findings demonstrate the practicality of CIL-based NIDS for real-world, heterogeneous IoT environments.
Detecting irregular patterns that suggest potential threats or system flaws plays a vital role in anomaly detection, which is essential for maintaining the security and integrity of network systems, especially in IoT environments where devices are often vulnerable and widely distributed. This study uses the UNSW-NB15 dataset, a vast collection of network traffic data, to examine several Machine Learning (ML) and Deep Learning (DL) approaches for network anomaly detection. In order to improve model performances, the study uses strategies such the Synthetic Minority Over-Sampling Technique (SMOTE) to address the issue of class imbalance. The capabilities of a number of widely used machine learning (ML) algorithms, such as Decision Trees, Random Forests, KNN, XGB, and widely used deep learning (DL) models, such as CNN, ANN and LSTM, to identify unusual patterns in various attack and typical behavior scenarios are assessed. Our findings provide important information for future advancements in anomaly detection techniques by highlighting the significance of feature selection, class balancing, and model resilience in successfully differentiating unusual behaviors in network traffic. XGB emerged as the most successful approach in this study, with its enhanced performance largely attributed to the synergy between its robust ensemble framework and the class balancing achieved through SMOTE.
Large Language Models (LLMs) have demonstrated high performance in text classification, particularly in specialized domains such as healthcare. However, their opacity raises concerns regarding interpretability, reliability, and governance. This paper explores integrating explainable artificial intelligence techniques with structured review to improve transparency and decision-making in LLM-based classification systems. We propose a framework that combines saliency-based methods, such as LIME and SHAP, with expert-in-the-loop validation to refine predictions and enhance interpretability. Through experiments on medical text classification, we study the effectiveness of integrating explainability with governance mechanisms. Results indicate that explainability-guided refinement improves classification accuracy while ensuring more interpretable and accountable outputs. This study provides insights into balancing performance and interpretability in high-stakes applications, supporting the adoption of LLMs in environments where transparency is critical.
In this paper, we present a novel AI-powered pipeline for estimating the weight of waste items using monocular images, designed for deployment in smart bins for automatic waste sorting. Developed by Ganiga.ai [1], a company focused on intelligent waste management, the system leverages a hybrid of zero-shot and depth estimation models to infer object volume and category, which are then mapped to weight estimates through material-specific priors. Our approach combines several state-of-the-art components: (i) a vision-language model (VLM) based on QWEN [13], distilled from multiple large language models (LLMs) including ChatGPT [17], Gemini [16], and DeepSeek [15] for label consensus and reasoning validation; (ii) zero-shot object detection via Grounding DINO [12]; and (iii) depth estimation using the Depth-Anything-V2 [14] model. The pipeline is calibrated with a 3D camera setup and integrated into cloud-connected bins distributed across Italy in public and commercial venues. We report results on a dataset of 1,000 real-world samples annotated with ground-truth weights. Experimental evaluation shows strong performance across waste categories (organic, plastic, glass, aluminium, paper, undifferentiated), with R2 scores up to 0.92, demonstrating the viability of vision-based weight estimation for scalable waste analytics and intelligent infrastructure.
The treatment and prognosis of a patient with breast cancer depend fundamentally on the TNM (classification of malignant tumors) staging, which requires a complex and delicate diagnostic process. This involves a combination of different techniques, including some invasive such as biopsies to assess lymph node involvement, and some non-invasive like imaging studies, including mammography, ultrasound, MRI, or CT scans, to determine tumor size and potential metastases. These procedures can be tedious for the patient and require significant clinical resources.This study explores the potential of hybrid PET/MRI imaging as a tool for simultaneous metabolic and morphological analysis, enabling precise breast cancer staging through a single test. PET/MRI images from 30 breast cancer patients were segmented in 3D to define regions of interest (ROI) in the breast. From these, 174 radiomic and textural features were extracted, capturing both morphological and metabolic tumor characteristics. These features were used to train multiple predictive models, including Fine Tree, Naive Bayes, SVM, KNN, and Bagged Tree, to evaluate their ability to determine the presence or absence of metastasis (M stage).The best-performing model employed 20 selected features and achieved an Area Under the Curve (AUC) of 93%, indicating high predictive accuracy. Evaluation metrics, such as the confusion matrix and AUC, highlighted the significant relevance of radiomic features derived from PET/MRI images for metastasis detection. This study underscores the potential of PET/MRI imaging in breast cancer diagnosis, offering a more streamlined and less invasive approach to TNM staging, which could improve patient outcomes and reduce diagnostic burden.
Monitoring food and detecting its freshness is crucial for ensuring food safety, reducing waste, and optimizing supply chains. Traditional methods rely on retroactive expiration dates, and recent models do not consider real-time temperature fluctuations, leading to inaccurate shelf-life estimation and unnecessary waste. This article presents an innovative framework for food monitoring and freshness scoring systems integrating the IoT and sensors. A unique freshness scoring model is developed to calculate the freshness score in numerical value and categorize the current state of the food into different classes based on its temperature. In addition, sensors and IoT devices are utilized to collect and analyze food temperature data. As IoT and sensors are integrated, this proposed system can continuously monitor temperature and calculate the freshness score at any stage of the supply chain. The result analysis shows that the IoT device can continuously collect the real-time temperature of food using the sensor. Then, the proposed system can effectively calculate the innovative freshness score using the proposed model. This proposed system can help businesses and consumers make informed decisions by providing real-time freshness insights and calculating the freshness score, ultimately improving safety and sustainability.
Chest X-ray imaging is crucial for verifying the correct placement of medical devices. However, deep learning models often fail to generalize across institutions due to domain shifts caused by variations in imaging protocols and annotation practices. This work addresses the challenge of classification of misplaced medical devices in plain chest X-ray images under domain shift by integrating domain adaptation into the training process. It proposes a framework that combines a Vision Transformer architecture with Deep Adaptation Networks, using the Maximum Mean Discrepancy loss to align the feature distributions between a large public dataset and a smaller, clinically distinct private dataset. Our results demonstrate that domain adaptation enables the detection of previously undetected misplacements in the target domain and enhances performance on the source domain, suggesting improved feature generalizability. The baseline model achieved high accuracy but failed to identify true positives in the target domain. In contrast, despite having lower precision, the adapted model increased the number of true positives from zero to twelve.
Road traffic surveillance is amongst the fields where artificial intelligence can play a major role in smart cities, where video and audio surveillance can be leveraged. Indeed, audio information is increasingly being used by surveillance systems to improve their effectiveness. This paper describes a novel method for detecting anomalous sound events in road traffic monitoring, such as car accidents and other hazardous events using audio data. To detect anomalies, the method combines generative variational autoencoders and interval type-2 fuzzy sets. The reconstruction error of each audio segment is computed using a baseline variational autoencoder, which offers a primary assessment of outlierness through thresholding. An interval type-2 fuzzy membership function with an optimistic/upper component and a pessimistic/lower component is employed to account for the uncertainty associated with this decision-making process. The final class attribution is made by interval comparison, based on a probabilistic technique. The evaluation results obtained after defuzzification reveal that the proposed membership function effectively enhances the performance of the baseline variational autoencoder.
Vulnerability scanning remains a crucial step of cyber defence, but at the same time, current tools face a trade-off between detection speed and the ability to uncover unknown threats. Signature-based methods provide fast verification of known exploits but fail to detect zero-day vulnerabilities, while fuzzing enables deeper exploration at the cost of time and resources. This paper introduces HyFuzz, a two-stage hybrid scanning framework that combines deterministic Common Vulnerabilities and Exposures (CVE) correlation with adaptive, AI-driven fuzzing. The first stage performs rapid fingerprinting and version matching using service banners, cross-referenced against a local CVE database and the Metasploit framework. If an exploit is confirmed, the scan terminates within approximately 3 minutes per host.If no CVE match is found, the second stage triggers coverage-guided fuzzing using BooFuzz or Hypothesis, followed by corpus expansion via a lightweight Generative Adversarial Network (GAN) or a zero-shot prompt of the 8-billion-parameter DeepSeek-r1 language model. Experiments on Apache, Nginx, and IIS demonstrate that this hybrid strategy achieves up to 10× faster detection times and a 35% improvement in crash discovery over baseline fuzzers.Although evaluation was conducted on HTTP (Hypertext Transfer Protocol), the framework is protocol-agnostic and applicable to Message Queuing Telemetry Transport (MQTT), Modbus, or Constrained Application Protocol (CoAP). The proposed "CVE-first, fuzzing-second" pipeline offers a scalable and efficient approach to vulnerability detection across diverse networked environments.
Misconfigurations within Operational Technology (OT) environments represent a significant source of cyber risk, often resulting in critical disruptions to industrial processes. However, the absence of standardized methodologies for quantifying their impact hinders effective risk assessment and prioritization. This study proposes a novel and fully automated framework that maps misconfigurations to the Common Weakness Enumeration (CWE) taxonomy through semantic similarity techniques, employing state-of-the-art sentence embedding models and cosine similarity metrics. The framework enables the computation of quantitative risk indicators by linking the identified CWEs to associated Common Vulnerabilities and Exposures (CVEs) and aggregating their Common Vulnerability Scoring System (CVSS) scores. A voting ensemble of pre-trained language models is introduced to enhance robustness and semantic accuracy. Experimental validation demonstrates improved precision over single-model baselines, confirming the efficacy of the proposed approach. The resulting system offers a scalable, data-driven tool for OT stakeholders to evaluate and prioritize misconfiguration-related cybersecurity threats systematically.
In the field of cybersecurity, the weakestlink principle creates inherent vulnerabilities, making it difficult to achieve absolute protection. Concurrently, the parallel continuous advancement in internet-technology-big data ecosystem has resulted in an ever-expanding attack surface, leading to the emergence of new-age threats. These include unknown attacks, which are complex and challenging to detect, due to irregular network behaviors and the absence of known attack signatures. The use of artificial intelligence(AI) techniques, particularly ensemble methods, has significantly enhanced this detection process. This study proposes a novel two-stage approach for accurately identifying unknown attacks. The first stage focuses on merging three real-world network traffic datasets. The second stage focuses on incorporating advanced and ensemble learning methods. Overall, the proposed methodology encompasses: (1) class imbalance handling through Synthetic Minority Over-Sampling Technique(SMOTE)and Adaptive Synthetic Sampling(ADASYN), (2) feature selection with Random Forest(RF) and Extreme Gradient Boosting(XGBoost), (3) attack detection using three boosting, one stacking, one bagging and three deep learning(DL) algorithms and (4) Hyperparameter optimization using Optuna. Following this approach, we have achieved promising results with high pass rates surpassing those of previous studies-including: 88-97% with RF+SMOTE and 78-82% with ADASYN+XGBoost.
In this paper, we analyze two advanced control strategies, fuzzy PI (Proportional-Integral) and fractional order Regulators (FOR), that are applied to a hybrid excited synchronous generator (HESG). Fuzzy logic's adaptability is utilized by the Fuzzy PI controller to handle non-linearities and uncertainties, while the Fractional Order Regulator, which has been optimized by Genetic Algorithm (GA), offers enhanced precision through non-integer order dynamics. The main objective is to measure and contrast the performance of these controllers in terms of overshoot, settling time and stability under different operational conditions. The GA-optimized fractional order regulator provides superior control performance, especially in scenarios requiring high precision and stability, as demonstrated by simulation results, which will reveal the strengths and limitations of each approach.
