OPTIMIZATION OF A FLAT-DIE PELLETIZING MACHINE FOR FISH FEED PRODUCTION USING MACHINE LEARNING MODELS

Abstract

High-quality fish feed production is a critical factor in the success of aquaculture systems. The quality of feed pellets, particularly their durability (pellet durability index), directly affects feed efficiency and aquatic environmental quality. This study applied machine learning approaches to predict and optimize pellet durability index in a flat die pelletizing machine used for fish feed production. The dataset consisted of 3,000 observations with 13 operational features collected through IoT sensors. Descriptive statistics were used to summarize overall data characteristics. The target variable was binarized to classify pellets as acceptable or non-acceptable. To address data imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied, followed by data cleaning and feature scaling to ensure uniformity. The preprocessed dataset was split into training and testing subsets. Six machine learning models were developed and evaluated using four performance metrics. Logistic Regression and Support Vector Machine achieved the best performance, with accuracy, recall, precision, and F1-score values of 0.756, 1.000, 0.756, and 0.861, respectively. The results indicate that operational factors in the fish feed pelletization process can effectively train predictive models to assess feed quality. The integration of machine learning in fish feed manufacturing offers potential improvements in production efficiency and feed quality for commercial aquaculture applications.