Sensor Data Modelling for Anomaly Detection in Aquatic Environments

Monitoring essential water parameters like pH, pond temperature, and dissolved oxygen is crucial for ensuring the production of high-quality aquariums. Machine learning techniques are under development to predict the periodic fluctuations of these factors, aiding fish cultivators in data-driven decision-making processes. Advanced real-time data collection, storage, and remote monitoring technologies facilitate the development of highly precise machine learning models. However, fish growers sometimes do not have access to sophisticated monitoring equipment and must instead use handheld tools for manual assessments. The analysis of data is limited in quantity and frequency due to the constraints of manual assessment methods. This study investigates the application of machine learning models like artificial neural networks, random forests, and multivariate linear regression to analyze water quality metrics in aquaculture systems with limited data. The study presents a modeling approach for estimating unobserved variables based on observed measurements and making predictions with limited training data in two scenarios. Our findings demonstrate accurate prediction of dissolved oxygen, pond temperature, pH, ammonia, and ammonium using random forest algorithms, even with water quality data measured only twice a day. Moreover, integrating these predictive models into a mobile device-accessible information system enables their implementation on smartphones, ensuring feasibility and cost-effectiveness.