This study addresses the growing need for intelligent monitoring in aquaponic systems by developing a predictive system based on artificial intelligence and environmental sensing. The goal is to improve fish welfare through the early detection of adverse water conditions. The system integrates low-cost digital sensors to continuously measure key physicochemical variables—pH, dissolved oxygen, and temperature—using these as inputs for real-time classification of fish health status. Four supervised machine learning models were evaluated: linear discriminant analysis (LDA), support vector machines (SVMs), neural networks (NNs), and random forest (RF). A dataset of 1823 instances was collected over eight months from a red tilapia aquaponic setup. The random forest model yielded the highest classification accuracy (99%), followed by NN (98%) and SVM (97%). LDA achieved 82% accuracy. Performance was validated using 5-fold cross-validation and label permutation tests to confirm model robustness. These results demonstrate that sensor-based predictive models can reliably detect early signs of fish stress or mortality, supporting the implementation of intelligent environmental monitoring and automation strategies in sustainable aquaponic production.

This study evaluates the reliability and suitability of long-term water quality data collected from a sustainable aquaponics system equipped with a pH, dissolved oxygen, and temperature regulation control strategy based on a PID algorithm. Although PID control was implemented to maintain parameters within optimal biological ranges, natural fluctuations and out-of-range measurements were recorded, particularly in pH. Rather than being considered anomalies, these deviations represent realistic environmental variations that must be captured for comprehensive system analysis. A thorough data validation process was conducted, including descriptive statistics, outlier detection, correlation analysis, principal component analysis (PCA), and temporal stability evaluation. Results confirmed the absence of missing data, the presence of controlled variability in dissolved oxygen and temperature, and meaningful correlations between parameters, with pH showing the highest variability. Autocorrelation and long-term trend analyses indicated stable measurement patterns that reflect real-world aquaponic dynamics. The validated dataset provides a robust foundation for future studies, particularly for the development of artificial intelligence (AI)-based predictive models aimed at early detection of fish distress or mortality.

Contexto: Los sensores electrónicos tienen un papel crucial en diferentes aplicaciones como la robótica o la automatización industrial o doméstica. Los sensores pueden medir variables ambientales esenciales para alimentar algoritmos digitales de procesamiento de señales y realizar acciones de una manera más eficiente. Los sensores que se utilizan para mediciones de distancia siguen enfoques diferentes. Sin embargo, es difícil encontrar un estudio con comparaciones de desempeño. Método: Se realizó un estudio empírico para evaluar y comparar el rendimiento de sensores ultrasónicos e infrarrojos en situaciones de detección frontal y lateral, considerando variaciones de ángulo y distancia. Resultados: Los resultados muestran que el sensor ultrasónico detecta la distancia con buena precisión a lo largo de su rango operacional. Sin embargo, la medida de distancia se torna inexacta cuando el obstáculo no está ortogonal al sensor. El sensor ultrasónico mostró gran precisión en la detección de largo alcance de obstáculos frontales, mientras que el sensor infrarrojo se desempeñó mejor en distancias cortas con obstáculos en ángulo. Un análisis estadístico confirmó la presencia de correlaciones lineales fuertes, especialmente en el caso del sensor ultrasónico, lo que respalda el uso complementario de ambos sensores en aplicaciones de medición de distancias. Conclusiones: La evaluación de sensores ultrasónicos e infrarrojos para la medición de distancias en aplicaciones que involucran robótica y el Internet de las cosas reveló que los primeros son más fiables para obstáculos distantes y ortogonales, mientras que los segundos funcionan mejor en distancias cortas y con superficies en ángulo, resaltando sus fortalezas complementarias y la necesidad de futuras mejoras para abordar la sensibilidad ambiental y otras limitaciones de detección.

This study presents the design and energy performance evaluation of a low-cost intelligent irrigation system tailored for cocoa (Theobroma cacao) agroforestry in tropical Colombia. The system integrates DHT11 temperature and humidity sensors, as well as analog soil hygrometers, with a Raspberry Pi 4 for edge computing. A multilayer perceptron (MLP) model, trained using Edge Impulse, was deployed on an Arduino Uno to enable real-time, autonomous irrigation control. Field validation was carried out over nine weeks in Piedecuesta, Colombia, with environmental variables recorded at 60-second intervals. Compared to conventional timer-based systems, the proposed solution reduced water consumption by 25%, while maintaining soil moisture consistently within the agronomic threshold of 70–85%. Post-irrigation measurements revealed stable microclimatic conditions, with relative humidity maintained between 81% and 87%. In terms of energy performance, the system operated at an average daily consumption of 204 Wh, powered entirely by a standalone solar photovoltaic unit. The deployed neural network model achieved 82% classification accuracy in predicting irrigation states based on sensor data. These results underscore the system’s potential as a scalable, energy-autonomous solution for smallholder agriculture in infrastructure-limited tropical settings.