Technological University of Pereira

Clima y Caf & eacute; [Climate and Coffee] supports Colombian coffee farmers' adaptation to climate change. The project finds and makes adaptation information available in accessible ways for farmers, while drawing attention to the crisis on a regional, national, and international scale. The website was built through iterative design testing with farmers and is targeted toward local associations that have internet access, are trusted information dissemination hubs, and are reported as central in conversations. Complementary forms of outreach redress internet connectivity challenges facing many farmers and include printed materials, radio segments, podcasts, WhatsApp videos, and in-person seminars. This field report reviews the project's development since its origins in 2018, providing a potential blueprint for other engaged research projects promoting environmental and agricultural outcomes. The report moves chronologically through the project's lifespan, beginning with the formative research that informed its design and moving through project development, external partnerships, leadership transitions, and change management.

This article introduces COFINFAD (Colombian Fintech Financial Analytics Dataset), a single-company 12-month dataset containing comprehensive behavioral information from 48,723 customers (representing the complete active customer base) of a Colombian fintech company. The dataset spans January 4 to December 29, 2023, capturing 3,159,157 transactions alongside demographic profiles, product usage patterns, customer satisfaction metrics, and digital app engagement behaviors. Data collection involved API integration with the company’s CRM system, transaction databases, and customer surveys (6,965 respondents, 14.3% response rate), while prioritizing customer privacy through multi-layered anonymization (pseudonymization, generalization, k-anonymity verification). The dataset comprises 57 variables including transaction frequency and value, product portfolio composition (savings, credit cards, loans, investments, insurance), customer satisfaction scores, Net Promoter Score (NPS), app login frequency, feature usage patterns, and support interaction metrics—all measured in Colombian Pesos (COP). COFINFAD is distinguished by its integrated multi-dimensional design, combining transactional, demographic, satisfaction, and digital engagement data unavailable in existing publicly available fintech datasets. The dataset is openly accessible via Mendeley Data (DOI: 10.17632/mhb4zn3258.1) under CC BY 4.0 license, adhering to FAIR (Findable, Accessible, Interoperable, Reusable) data principles.

This paper proposes a horizon-one nonlinear model predictive control (MPC) strategy for permanent magnet synchronous motors that combines formal stability guarantees with real-time implementability. The controller is derived directly from the original bilinear discrete-time model, avoiding linear approximations commonly used in motor-drive applications. Despite the nonlinear formulation, the associated optimal control problem admits a closed-form analytical solution with a unique global optimum, enabling efficient implementation on resource-constrained microcontrollers. Adaptivity is achieved through a recursive nonlinear identification scheme and a disturbance observer, which allow online estimation of uncertain parameters and mechanical torque without direct torque measurements. Closed-loop stability is established using discrete-time Lyapunov theory, and sufficient conditions for asymptotic stability and region-of-attraction estimation are provided. Simulation and experimental results demonstrate that the proposed approach achieves improved transient performance and reduced overshoot compared to conventional field-oriented control, while maintaining low computational cost and robust operation under disturbances.

The detection of titanium dioxide (TiO2) nanoparticles is a significant challenge due to their extensive industrial use and potential health and environmental impacts, which demand accurate, label-free approaches. This work presents an automatic detection system based on spectroscopy with optical frequency combs (OFC) in dual-coupled microresonators. The OFC generation was modeled through the Lugiato-Lefever equation, while propagation in distilled water containing TiO2 was simulated using the finite element method (FEM). The water–TiO2 mixture was described with the Yamaguchi model in a 5 × 5 mesh to represent non-uniform concentrations. From the norm of the electric field at a probe, a database of 11 classes (0–100%) with controlled Gaussian noise was constructed. A Transformer-based classifier was trained and compared with 1D-CNN and SVM under Monte Carlo validation (100 random 70/30 splits). The Transformer achieved 99.84 ± 0.01% accuracy with an inference time of 0.793 ± 0.05 s, while the 1D-CNN reached 99.64 ± 0.09% and the SVM 84.73 ± 1.48%. A repeatability test with 200 iterations confirmed deterministic DKS trajectories. The results demonstrate that combining dual-coupled microresonators, FEM, and Transformer architectures enables precise and efficient detection of TiO2 nanoparticles in aqueous solutions.

Obtaining high-quality labeled data for supervised learning is costly, motivating the use of crowdsourcing, which distributes the annotation process across multiple workers with varying levels of expertise. A key challenge in crowdsourced data is annotation sparsity, as each worker labels only a limited subset of instances. This sparsity can amplify class imbalance, reduce supervision for minority classes, and bias standard cross-entropy-based models toward the majority classes. To address this problem, we propose a correlated chained Gaussian process framework trained on a focal-loss-based variational objective (CCGPFL). This probabilistic framework jointly models latent ground-truth and instance-dependent annotator reliability while accounting for correlations among annotators. In addition, the focal-weighted objective mitigates the imbalance induced by sparse annotations by assigning greater importance to harder examples during training. Experiments on synthetic, semi-synthetic, and fully real multi-annotator datasets show that CCGPFL achieves competitive and often superior performance relative to state-of-the-art learning-from-crowds baselines in terms of Overall Accuracy (OA) and Area Under the ROC Curve (AUC).