Validación de Soft Sensing con Redes Neuronales e Híbridas y Prescripción en Sistemas de Decisión para la Gestión del Riesgo Nitrogenado en Acuicultura

Abstract

Intensive tilapia aquaculture faces toxicity risks from nitrogenous compounds, whose nonlinear behavior makes them difficult to predict with conventional monitoring. To bridge the gap between water‑quality diagnosis and on‑farm action, this study validates a neural soft sensor and a prescriptive Decision Support System (DSS) as an integrated architecture for nitrogen‑risk management. Methodologically, a dataset of 291.812 records was consolidated by combining open observational sources with synthetic scenarios of critical toxicity. Five machine‑learning architectures were evaluated on this corpus, and a deterministic DSS based on biological rules for Oreochromis spp. was designed. The results show that a multilayer perceptron (MLP) achieved a macro‑F1 of 0.9964 and a mean latency of 0.53 ms, an ideal range for IoT deployments. These metrics are interpreted as a validation of computational feasibility under a labelling scheme derived from deterministic rules, rather than as a guarantee of immediate generalization to field conditions. When integrated into the DSS, this neural core achieved a zero under‑prescription rate in high‑risk scenarios and a conservative over‑prescription rate of about 60%, supporting a safety‑first design policy that avoids omitting critical interventions. The prototype provides a solid methodological reference for future intelligent systems in regional aquaculture.