Causal Analysis of Intelligent On-Demand Fish Feeding, Feeding Regimes, and Water Quality in Aquaculture Using DoWhy and Time Series Data From an AIoT System

Efficient feeding and water quality (WQ) management are critical to sustainable aquaculture, yet many existing approaches rely on correlation-based analyses that overlook the true causal relationships among environmental and operational variables. This study applies the DoWhy causal inference framework to real-time time series data from an AIoT-enabled aquaculture system to uncover how intelligent on-demand fish feeding, feeding regimes, and WQ interact causally. Over a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$10$</tex>− day trial in a tilapia pond, an intelligent fuzzy-based decision support system (DSS) governed feeding based on sensor inputs, trigger timing, and computer vision outputs, while WQ datatemperature (Temp), pH, dissolved oxygen (DO), and total dissolved solids (TDS)-were logged continuously via IoT sensors. Causal analysis incorporated correlation assessment, the PC and GES algorithms, ICA-LiNGAM, and backdoor adjustment, with robustness verified through refutation tests. Results identified Temp, pH, and feeding regime as key drivers: Temp had a strong positive causal effect on both TDS (ATE = 71.389, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$p=0.002$</tex>) and feeding regime (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{ATE}=0.339, p=0.012$</tex>); <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{p H}$</tex> negatively impacted time, day number, and Temp, but positively affected DO; and feeding regime significantly influenced feed triggers (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{A T E} \boldsymbol{=} \mathbf{1 6. 2 6 3}, p=0.071$</tex>). The stability of marginal effects under refutation highlights the reliability of these insights. Overall, this work demonstrates how causal modeling can inform more adaptive, interpretable, and efficient control strategies in intelligent aquaculture systems.