Optimising maize threshing process with temporal proximity soft actor-critic deep reinforcement learning algorithm

Maize threshing is a crucial process in grain production, and optimising it is essential for reducing post-harvest losses. This study proposes a model-based temporal proximity soft actor-critic (TP-SAC) algorithm to optimise the maize threshing process in the threshing drum. The proposed approach employs an LSTM model as a real-time predictor of threshing quality, achieving an R2 of 97.17% and 98.43% for damage and unthreshed rates on the validation set. In actual threshing experiments, the LSTM model demonstrates an average error of 5.45% and 3.83% for damage and unthreshed rates. The LSTM model is integrated with the TP-SAC algorithm, acting as the environment with which the TP-SAC interacts, enabling efficient training with limited real-world data. The TPSAC algorithm addresses the temporal correlation in the threshing process by incorporating temporal proximity sampling into the SAC algorithm's experience replay mechanism. TP-SAC outperforms the standard SAC algorithm in the simulated environment, demonstrating better sample efficiency and faster convergence. When deployed in actual threshing operations, the TP-SAC algorithm reduces the damage rate by an average of 0.91% across different feed rates compared to constant control. The proposed TP-SAC algorithm offers a novel and practical approach to optimising the maize threshing process, enhancing threshing quality.