# Quantification of feeding intensity and feeding control of largemouth bass based on water surface vibration characteristics

## TL;DR

This study uses water surface vibrations and deep learning to precisely control feeding in largemouth bass aquaculture.

## Contribution

A novel vibration-based model using LSTM for real-time feeding control in aquaculture is developed and validated.

## Key findings

- Vibration displacement increases significantly with fish size and density.
- The LSTM model outperformed GRU and Transformer in predicting feeding intensity.
- The vibration-LSTM system achieved low residual feed rates and lower costs than other methods.

## Abstract

In response to the demand for precise feeding in high-density aquaculture, this study established a dynamic prediction model for fish feeding intensity by integrating vibration signal quantification and deep learning. Through multidimensional experiments (fish size: 50–300 g; stocking density: 20–60 fish/group; feeding speed: 1-3 g/s; feed particle size: 2#4#6#), we quantified the three-axis displacement signals of Micropterus salmoides during feeding. Results demonstrated significant effects of all parameters on water surface fluctuations (p < 0.05). Vibration displacement exhibited linear relationships with fish size and density. The 300 g group showed 109.7% higher peak amplitude than the 50 g group, while the 60-fish density group exceeded the 20-fish group by 141.9%. Optimal palatability (4#) reduced fluctuation frequency by 42%. A predictive model for feeding vibration patterns was developed, incorporating fish size (S), density (D), feeding speed (V), feed particle size (Φ), real-time triaxial vibration sum, and time series (t) as inputs to predict the summed vibration displacement at t + 5 s, which serves as a quantitative proxy for feeding intensity. The Long Short-Term Memory (LSTM) model accurately captured fish feeding dynamics (RMSE = 69.43 μm, MAE = 48.00 μm, R2 = 0.883). In comparative analysis, the LSTM outperformed Gated Recurrent Unit (GRU) and Transformer models in forecasting accuracy. Deployed on an embedded system (Orange Pi AiPRO), closed-loop tests demonstrated superior performance: residual feed rates were ≤ 0.8% across all trials, outperforming optical flow (2.69% residuals) and graph neural network (6.58% residuals) methods. The space complexity of the vibration-LSTM approach was only 6.4–31.8% of GCN-based approaches, enabling cost-effective (<$200) real-time control.