Defining effective policies for managing traffic in large cities, particularly near major logistics hubs such as ports, is a challenging problem due to the critical interaction between mobility and freight flows. In this work, we combine a traffic simulator with a change-detection test to assess a priori whether a specific policy would have an impact on city mobility. More specifically, we propose a general methodology to identify the expected number of days/monitoring samples before gaining evidence that the policy has introduced a detectable change in the traffic data acquired after enforcing the policy. Our experiments, conducted on simulated traffic, focused on the port-city context of Genova, showcase that our proposed methodology can provide outcomes that are consistent with the kind and expected effectiveness of policies under evaluation.
Large language models (LLMs) have a significant impact on the development of applied solutions for industrial purposes, with their general knowledge capabilities supporting more autonomous work processes. The weights learned from these extensively trained models provide the backbone for fine-tuning and retrieval-augmented generation (RAG) of LLMs, allowing them to address specific work-task operations in unique enterprise contexts. In this study, a locally running vision-based LLM framework is proposed, including custom vision-based RAG and a business-oriented evaluation method for handling sensitive core business documents. This approach enhances data privacy and prevents leakage of sensitive data. The LLM adapts core enterprise knowledge from in-house data spaces using affordable computational resources. Our proposed norm-chatbot for enterprise LLM data processing runs on a commercial Nvidia RTX 4070Ti GPU processor in an offline environment. A RAG vision technique is applied to handle complex document layouts by transforming them into image content, addressing the challenges that traditional text-based indexing and RAG methods face when documents consist of mixed data types such as text, tables, images illustrations, graphs, etc. The proposed framework uses RAG data processing with a PDF-to-image converter on top of an LLM vision model that embeds the images with textual information. The framework combines ColPali RAG tested with the LLAMA 3.2 11b Vision and LLAVA 1.6 models for highly semantic document processing, specifically for both descriptive and question answering (Q&A) tasks. From the study we found a clear trade-off between fast content generation and precision. However, the proposed framework shows promising results in generating contextual meaningful content in two different knowledge generation tasks for processing custom datasets of enterprise norms and a standard database to support tool development and operations in a Danish tool manufacturing company with a global customer segment.
Time series forecasting for sales is well established in the business world. However, the nature of the data and often the domain area of the application can yield unexpected inconsistencies and constraints. Machine learning tools can be utilized to enhance the accuracy of predictions and support effective forecasting. This work presents the case of a small chain of bakeries where machine learning supported forecasting aims to tackle food wastage, enhance profit margins, and contribute to a more sustainable community.
Dissolved oxygen (DO) is the amount of O-2 that is present in water and it is found in molecular state. Water bodies receive oxygen from the atmosphere and from aquatic plants. Running water, such as that of a swift moving stream, dissolves more oxygen than the still water of a pond or lake. Dissolved O-2 levels change according to the temperature, air pressure and salinity. It is consumed not only by the reduction of substances such as sulphides, nitrites and iron ions, but also by the respiration of microorganisms and the oxidative decomposition of organic substances by aerobic microorganisms. The aim of this research is the analysis and classification of DO at the Hydrological Station G. Giannouli (HSG) ofthe Pinios river in central Greece, with the development of robust Machine Learning models. The results clearly demonstrate that these models can successfully assign five labels based on the thresholds established by the Norwegian Water Research Institute namely: "Poor", "Deficient", "Moderate", "Good" and "High".
Digital twins have emerged as a powerful tool for industrial process monitoring and optimization, requiring accurate and computationally efficient models to capture system dynamics. This paper presents the implementation of Echo State Networks (ESNs) for modelling complex industrial systems with non-linear behavior. The study is conducted on a pilot plant representing an industrial process, where four variables (level, pressure, flow, and temperature) are controlled and estimated using ESN models. To assess the impact of different architectures, each model is tested with various configurations, including deep networks, feedback integration, and a combination of both. The results demonstrate that the effectiveness of these architectures depends on the dynamic characteristics of each variable. The most notable improvement is observed in process temperature estimation, where feedback significantly enhances performance. The results highlight that the combination of both, feedback mechanisms and deeper architectures, can improve the prediction of variables with slower dynamics and higher inertia.
Amyotrophic lateral sclerosis (ALS), a progressive neurological disorder impacting motor neurons crucial for voluntary muscle movement, poses significant challenges for affected individuals. In Pakistan, an estimated 3.5 per 100,000 people grapple with this debilitating condition. The primary focus of our project is to address the communication difficulties faced by ALS patients due to muscle weakness and loss of motor function. We propose the development of an intuitive application that integrates eye-gaze technology with a comprehensive "Phrase Bank", complemented by voice output capabilities. This innovative approach allows patients to select phrases and images by simply gazing at them on a screen, significantly reducing the physical effort required for communication, particularly as ALS progresses. Recognizing the importance of familiar phrases in everyday communication, our application includes a dynamic "Phrase Bank". This feature acts as a readily accessible repository of commonly used phrases and sentences, offering patients a collection of vital expressions to convey their needs, emotions, and thoughts. Navigating through this collection becomes an effortless task for users, ensuring the application remains both intuitive and efficient. While the precise cause of ALS remains elusive, with genetics believed to play a role in some cases, our project takes a proactive approach, focusing on improving the lives of individuals with ALS by providing innovative tools to enhance communication. This aims to enable a more fulfilling and connected life despite the challenges posed by the disease.
This paper presents a Decision Support System (DSS) for Zero Defect Manufacturing (ZDM), deployed in three factories. The DSS comprises modular components that enhance decision-making through advanced visualization for monitoring, prediction, analysis, and defect detection. Its design prioritizes usability, offering comprehensive visual analytics, easy management of connected devices, human-Artificial Intelligence (AI) interaction, and interface expandability. The DSS captures product images using RGB cameras, which are analyzed by AI algorithms to detect defects. Operators can review the detection results through a user-friendly web interface and provide feedback. The significant novelty of the system lies in its application of human-in-the-loop (HITL) Learning, where human feedback is used to retrain the AI models, progressively improving the accuracy of the defect detection, as demonstrated in real-world experiments. At the same time, operators learn from their interactions with the AI, while their empirical knowledge is complemented by real-time informed data. The system fosters a synergistic collaboration between human operators and AI, reducing defective products and inspection time. In one factory, the DSS led to 90
Landslides pose significant risks to infrastructure, ecosystems, and human lives, making accurate prediction crucial for disaster preparedness and mitigation. We integrate multimodal environmental data to enhance landslide prediction using machine learning. Specifically, we combine temporal weather data from ASOS, static vegetation data from NLCD, static soil composition data from SOLUS100, and temporal soil attributes from ERA5-Land to estimate landslide probability within a 5 km radius of ASOS weather stations across six U.S. states. We frame this as a multiclass classification problem, predicting high, low, or no landslide probability. Given the inherent imbalance in landslide occurrence, we explore various techniques such as SMOTE oversampling, class-weighted training, and dimensionality reduction to improve model performance. Our results indicate that XGBoost trained on SMOTE-balanced, PCA-reduced data incorporating all four datasets achieves the highest macro F1-score of 0.70. Analysis of feature importance reveals that significant predictors span all datasets, highlighting the necessity of integrating diverse environmental variables. Additionally, we conduct state-wise and seasonal comparisons to assess regional variations in model effectiveness. This research demonstrates the potential of multimodal data fusion and machine learning in landslide forecasting, paving the way for more robust and interpretable predictive models for natural hazard assessment.
The application of Machine Learning (ML) to Computational Fluid Dynamics (CFD) has gained significant attention due to its potential in speeding up simulations and approximating numerical solutions of physical equations. Still, the dominant role of ML, which is to infer expert labels that cannot be calculated from explicit equations, has received much less attention in CFD. A major challenge in this direction is the scarcity of large, annotated datasets required to train robust ML models for flow field classification. In this work, we address the problem of training a ML model to classify CFD flow fields, inferring pathologies affecting the human upper airways. We propose a novel data augmentation method to address this limitation which involves an automated pipeline to extract CFD-ready surfaces from CT scans and a computational geometry technique to generate synthetic training samples. By defining deformation functions for specific pathologies on a reference surface and mapping these to healthy anatomical surfaces, we create a large and diverse training set with minimal expert supervision. This method allows for the generation of a dataset with high anatomical variability and well-defined labels, enhancing the model's ability to generalize to unseen geometries. We demonstrate that a Neural Network (NN) can accurately classify two common nasal pathologies, septal deviation and turbinate hypertrophy, achieving strong performance on real pathological patient data despite being trained solely on synthetic samples.
Efficient water management is essential in regions facing persistent imbalances between water supply and demand. The Segura Hydrographic Confederation has developed a digital twin platform integrating Artificial Intelligence (AI) tools for efficient water resource management. This study presents an AI-based water quality monitoring module integrated in this platform, that explores three modeling approaches: ML-based water quality predictions, DL-based forecasting, and surrogate modeling. The study demonstrates that AI-enhanced water quality monitoring can support decision-making processes.
Boiling Liquid Expanding Vapour Explosion (BLEVE) is a high-energy explosion that generates intense blast loads, posing significant safety risks to infrastructure and public safety. Accurate overpressure prediction is essential for developing effective safety measures and emergency response strategies. Traditional empirical models fail to capture complex nonlinear interactions, while Computational Fluid Dynamics (CFD) methods, though accurate, are computationally expensive and impractical for large-scale and real-time applications. This study develops a neural network-based machine learning model for BLEVE overpressure prediction, offering an efficient alternative to traditional approaches. To improve model interpretability and reliability, Shapley Additive Explanations (SHAP) is employed, identifying key features that influence BLEVE overpressure. Using SHAP insights, an optimized model is constructed by selecting the most significant features, enhancing both accuracy and generalizability. The proposed approach is validated on both FLACS simulation data and experimental datasets, achieving a Mean Absolute Percentage Error (MAPE) of 2.93
Traffic congestion remains one of the most significant issues affecting highways worldwide. This problem is directly associated with productivity loss (due to time wasted in traffic jams), environmental pollution, increased fuel consumption, and adverse effects on human health. Toll booths constitute a major contributor to congestion, especially during peak hours and periods of high travel demand, such as holidays and vacation seasons. In this research, the authors examine the case of the Egnatia Odos Motorway, focusing on the Mesti-Komotini toll station over a three-year period. Using a large set of Nonlinear Autoregressive Neural Networks (NARNNs), the study forecasts vehicle crossing volumes based on past traffic values. The models were trained using the Hold-Out Validation Method, and their performance was evaluated using the R-squared (R-2) index, achieving values above 0.94. The results are highly promising, indicating that the more historical data is available, the higher the prediction accuracy becomes. The proposed models can serve as valuable tools for traffic management authorities to alleviate motorway congestion when necessary.
Floods rank among the most destructive natural disasters, inflicting severe harm on human lives, societies, and ecosystems. A flood can be characterized as a complex nonlinear phenomenon, thus the use of Machine Learning (ML) techniques and of Artificial Intelligence (AI) tools tend to be of vital importance. The authors of this paper propose an innovative hybrid approach that employs the Fuzzy C-Means (FCM) Algorithm and a novel Fuzzy Inference System (FIS) in order to determine the relation between the inflow and the outflow. Overall, the whole effort is based on a fuzzy rule-based methodology. It is a novel and hybrid research effort, that employs both FCM (for the determination of clusters) and a novel FIS for the outflow prediction. In fact, it incorporates fuzziness blurring the crisp margins of the clusters, offering a rational model. The model has proved to be quite robust in the testing phase, achieving an overall R-2 of 0.9104. Furthermore, the proposed hybrid approach seems capable of assessing the beginning of flood events.
Obstacle avoidance in aerial robotics remains a critical challenge, particularly in environments with uncertain terrain and weather conditions. This study introduces a Constrained Q-Learning model that leverages spatio-temporal data from LiDAR and OctoMap to achieve Zero-Shot Execution (ZSE) for autonomous Unmanned Aerial Vehicle (UAV) navigation in unseen environments, eliminating the need for iterative training. Experimental evaluations are conducted using a high-fidelity simulator across three environments: random forests, clustered forests, and metropolitan areas, under varying obstacle densities and flight velocities. The proposed model demonstrates a 100
This paper focuses on the problem of automatically forecasting mortality rates over long time horizons. In the context of the Lee-Carter model, an approach based on general regression neural networks is presented and discussed. Our proposal preserves the LC defining parameters and structure, adds flexibility at reduced costs in terms of complexity, and requires a weak human intervention for the identification of the optimal parameters. Moreover, GRNN models need relatively few data to train, an advantage useful in actuarial data. An application to real data shows that an additive GRNN model has, in general, better forecasting performances than both the multiplicative GRNN model and the KNN model, taken as a benchmark. Furthermore, between two different long term forecasting strategies, the analysis highlights how, in general, MIMO is preferable to the classic recursive procedure.
The COVID-19 pandemic has been a shacking experience for the entire world. Retrospective analysis of the data gathered during the pandemic can be used for the preparation for future pandemics. It is now possible to ascertain if the pandemic response has been the same across the world. Detecting differences in responses over time can be useful for preparation for future pandemics by researching on the causes for different responses. In this direction, this paper contributes evidence that the pandemic response, as measured by the death time series of each country, was not the same everywhere. Clusters of countries with similar death time series can be detected, such that countries in different clusters have rather different patterns of deaths. Dynamic Time Warping (DTW) allows the elastic matching of time series. The cost of DTW matching provides a measure of similarity between the time series. Applying Hierarchical Clustering it is possible to find these clusters. Our findings confirm the existence of a robust cluster of western Europe countries previously reported following rather diverse approaches. Furthermore, we examine in detail the different patterns of several representative countries relative to Spain.
Weather-related power disruptions present significant challenges to public infrastructure, societal well-being, and the distribution grid. Predicting outage durations in distribution grids is another challenge compared to transmission line outage durations due to distribution networks' complexity and finer granularity. While forecasting forced power outages is crucial, accurately estimating their duration is essential for timely response and mitigation measures. This study introduces the Spatiotemporal Multiplex Network (SMN-WVF), a methodology designed to predict power outage durations across varying lead times, tackling the difficulties posed by small, high-complexity spaces within distribution grids. SMN-WVF employs multiplex networks that incorporate multi-modal data across both time and space, including layers such as power outages, weather conditions, weather forecasts, vegetation, and distances between substations. We demonstrate the importance of incorporating additional layers of data sources as they are shown to help the model's predictions through gradual improvement in the macro F1 score performance.
In particle accelerator physics, superconducting magnets are crucial to achieve high-energy particle beams used in fundamental physics experiments. One of the key challenges of superconducting magnets is the transition of the superconducting material to the normal conducting resistive state-called quench-which could cause damage to the superconducting magnet volume, due to the large quantities of heat produced by ohmic losses in the material if the magnet is not properly protected. In this work, we train explainable machine learning models to detect the occurrence of quenches, starting from the harmonic decomposition of the magnetic field produced by the residual superconducting magnetization after a quench event and aiming at explaining the prediction process. A successful solution to this problem is a fundamental ingredient in the construction of a real-time predictive maintenance system. Indeed, interpretability paves the way to the construction of a tailored, efficient, and affordable system that only considers the relevant magnetic field harmonics. We show that quenches can be detected with high accuracy by rather simple models, and we also describe some preliminary but encouraging results on the quench localization problem.
Time series data continuously generated during CNC machining processes contain essential information on their performance and efficiency, but also on produced parts. Using industrial edge devices, among low frequency and event data also high frequency data at 1 kHz can be generated. Within these large, multi-dimensional datasets, interrelationships and semantics are usually not available without additional processing. Among others, pattern recognition algorithms have been used to find repeated patterns within the data or to find areas with constant technological parameters. However, the approaches often lack reference to parts and part characteristics, leading to individual analytics solutions. To face this issue, a data-driven, multi-level segmentation is developed. The first part of the algorithm extracts production processes of parts of the same product type; the second part segments these production process data by each individual part. For the second method, a discrete rule-based pattern recognition algorithm is proposed to perform the segmentation. Before, the available data are contextualized to gain additional process knowledge, enabling the effective use of the implicitly available information. This part-focused segmentation is verified on milling processes and not only allows for more effective analyses of CNC machining processes, but also enables the analysis of historical data which increases their value significantly.
Forecasting transportation demand is essential for the effective planning and operation of infrastructure and related services. Accurate predictions are critical to ensuring the efficient performance of transport systems and meeting evolving user needs. This study analyzes the key factors influencing transportation demand and explores their implications for infrastructure planning and management. The primary objective is to integrate Singular Spectrum Analysis (SSA) with Artificial Neural Networks (ANNs) for time series modeling and forecasting of transportation infrastructure demand. SSA, a non-parametric method for decomposing time series into lower-dimensional interpretable components, is employed to extract significant features such as trend, periodicities, and noise. This hybrid approach enhances model predictive accuracy while supporting efficient handling of large datasets. The methodology is applied to a real-world case study on the Egnatia Odos Motorway, focusing on the Iasmos–Komotini toll station over a four-year period. The models are trained using the hold-out validation method and evaluated based on the coefficient of determination (R2) and Mean Squared Error (MSE).
