Abstracts Track 2026

Area 1 - Data Engineering, Infrastructure and Business Applications
Area 2 - Data Science & Machine Learning

Area 1 - Data Engineering, Infrastructure and Business Applications

Nr: 243
Title:

Real-Time Data Pipeline for Smart Irrigation Systems
Authors:

Carlos Lemonde, Teresa Afonso do Paço, Jucilene Siqueira, Livia Pian, Raúl Estevinha and João Rolim
Abstract: Water scarcity is one of the main global challenges in the 21st century, driven by increasing demands on freshwater resources for agriculture. An efficient irrigation system requires an accurate management to avoid using more water than necessary, increasing water productivity and reducing energy pumping costs. Optimization of irrigation systems does not rely solely on precise control, but also involves continuous real-time monitoring of crop conditions. Deployed sensors across the field measure soil, plant and environmental variables, generating vast amounts of data collected daily. The analysis of such data enables actionable insights towards enhanced agriculture performance. This work proposes a data pipeline architecture that can be applied to olive orchard irrigation, describing every step in the process of knowledge discovery. From the extraction and integration of sensor data, through processing, to visualization and analysis. It aims to support decision-making by inferring plant water requirements via the estimation of tree transpiration based on field data, while uncovering meteorological, physiological and soil hidden patterns with the aid of machine learning models.

Area 2 - Data Science & Machine Learning

Nr: 208
Title:

Industrial Applications of Machine Learning for Real-time Root-Cause Analysis and Process Improvement
Authors:

Timothy Mark Young and Donald Hodges
Abstract: The fields of machine learning (ML) and artificial intelligence (AI) are evolving at an unprecedented pace, fundamentally reshaping industries worldwide. In manufacturing, AI presents significant opportunities to improve system performance by enabling advanced problem-solving approaches that reduce process variation and drive substantial cost savings. This study focuses on the application of ML algorithms, specifically gradient-boosted trees and random forests, within a real-time manufacturing environment. The facility under study produces bio-based composite panels. A high-quality, integrated database was available, containing 463 process variables and three dependent variables related to final product strength properties. These process variables included parameters such as pressing temperatures, drying conditions, line speed, moisture content, and forming weight. The ML analysis identified three key process variables with significant influence on strength properties: ‘Dwell time in digesting’; ‘Refining input auger speed’; and ‘Press temperature’. Among these, ‘Dwell time in digesting’ was consistently ranked as the most influential variable across models. A notable outcome of the study was the short time period of discovery and validation of ML results. Based on these findings, an adjustment to the ‘Dwell time in digesting’ was implemented, resulting in an increase in average tensile strength of approximately 0.5 N/mm². This improvement was validated through additional testing within 24 hours, demonstrating both the robustness and operational value of the approach. The increase in tensile strength creates opportunities for significant cost savings, either through increased line speeds or reduced adhesive usage, without compromising product quality.
Nr: 242
Title:

LLM-Based Semantic Structuring for Clustering Noisy Civil Complaint Texts
Authors:

Shinmi Ahn, Hangyul Kim and Nayoung Kim
Abstract: The Korean Safety e-Report system handles various non-emergency safety-related civil complaints, including safety hazards, daily inconveniences, illegal parking, and traffic violations. The system accumulates large-scale user-generated complaint data; however, automatic analysis and classification remain challenging because complaint texts often contain noisy characteristics such as colloquial expressions, typographical errors, and incomplete sentence structures. This study analyzes the impact of hybrid semantic representation, which combines raw complaint texts with LLM-extracted object, status, and location information, on text clustering performance. Using 19,955 complaint records collected from the Korean Safety e-Report system in 2024, structured semantic information was generated using the Qwen2.5-7B-Instruct model. Clustering performance was then compared between raw-text representations and hybrid representations using various embedding models (TF-IDF, Ko-SBERT, BGE-M3, and Multilingual E5) and clustering algorithms (K-means, Agglomerative Clustering, HDBSCAN, and Spectral Clustering), evaluated with Silhouette Score (SS) and Calinski-Harabasz Index (CHI). Experimental results showed that the combination of BGE-M3 and HDBSCAN achieved the best performance, improving the SS from 0.567 to 0.579. TF-IDF-based experiments also showed performance improvements, indicating that LLM-derived semantic features can compensate for the limitations of sparse representations. Ko-SBERT showed limited performance improvements, whereas E5 showed performance degradation in some clustering settings. These findings suggest that compatibility between structured semantic information and embedding models is an important factor in clustering civil complaint texts and demonstrate the applicability of LLM-based semantic structuring in public complaint systems.