Soft Sensing Measurement of Dissolved Ammonia Nitrogen in Tank-Based Eel Aquaculture Systems Utilizing Deep Learning

Japanese eels are susceptible to low concentrations of dissolved ammonia nitrogen $(\mathrm{N}\mathrm{H}_{3}-\mathrm{N})$. To provide proper intervention, measuring the dissolved ammonia nitrogen in the tank-based eel aquaculture is crucial. However, existing laboratory methods and digital sensors are timeconsuming and expensive solutions. This study proposed an Attentive Interpretable Tabular Learning (TabNet) Deep Learning soft-sensing model to predict dissolved NH3-N concentrations. The prediction model is based on input parameters: water temperature, conductivity, pH, and dissolved oxygen. The datasets are gathered from the Global Environment Monitoring System (GEMStat), a water quality parameter database. The performance of the proposed Deep Learning model was compared with Feed-forward neural network (F-NN) and One Dimensional-CNN. The models were evaluated using statistical methods: R-squared (R2), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE). The simulation results showed that TabNet performed better in predicting dissolved NH3-N concentrations.