A prediction method for remaining feed weight in intelligent feeders from electrical signal time series

Feed interruptions in intelligent feeders seriously disrupt production continuity, leading to insufficient feed intake and triggering competitive feeding behavior among pigs. In current research on intelligent feeders, the monitoring of remaining feed predominantly relies on manual methods, indicating a lack of more effective automated monitoring approaches. However, the electrical signals generated during equipment operation possess rich representational capabilities and have been widely applied to guide production. Inspired by this, a two-stage remaining feed weight prediction model based on voltage and current signals is constructed in this study. The model leverages the ability of wavelet transform to simultaneously extract time-domain and frequency-domain features from electrical signals, enabling it to capture both short-term fluctuations and long-term trends. In the first stage, wavelet features are used to predict the feed fines ratio, which serves as an input for the subsequent stage, significantly enhancing the accuracy of remaining feed weight prediction. In the second stage, a transformer-based model incorporating time–frequency attention mechanisms and uncertainty analysis is employed to output the remaining feed weight and its corresponding confidence interval. This study quantifies the contribution of different wavelet features to the prediction results by performing wavelet feature importance ranking and SHAP mean value ranking, thereby enhancing the model's interpretability. Additionally, uncertainty analysis was conducted on the model. The results demonstrate a high agreement between the predicted intervals and actual values, and a significant positive correlation was observed between prediction uncertainty and actual errors. Comparative experiments show that the WaveResNet outperforms both DNN and ResNet models, achieving an MAE of 1.4950, R M S E of 2.6467, and an R 2 of 0.9296. The WaveTransformer model outperforms conventional deep learning models—including RNN, LSTM, GRU, TCN, and Transformer—achieving an MAE of 1.6663, an R M S E of 2.3719, and an R 2 of 0.9574. The model demonstrates lightweight architecture and low latency during practical deployment, effectively meeting the real-time requirements of pig farm applications. • A novel two-stage model named WRN-WTF is proposed for predicting the remaining feed weight in intelligent feeders. • The Wavelet transform is applied to decompose electrical signals and extract time–frequency features. • The feed fines ratio is used as an input feature to enhance prediction accuracy. • Uncertainty analysis is incorporated to quantify the model's predictive confidence. • WRN-WTF outperforms similar models and has significant practical applicability.