A prediction approach of unthreshing grain loss for wheat harvest employing an improved BiLSTM method

• A BiLSTM-based model is proposed to predict the wheat unthreshed grain loss. • Temporal feature fusion optimizes the representation of multi-parameter coupling. • Integration design of GCNN, ResNet and TA enhances feature extraction ability. • This approach enables precise prediction of wheat unthreshed grain loss. The complex working environment and multi-parameter coupled non-linear process during wheat combined harvesting pose great challenges for predicting harvesting losses. To accurately obtain the evolution trend of unthreshing grain losses and implement loss suppression control in advance, this paper proposes a BiLSTM-based prediction model, integrating a Gaussian convolutional neural network (GCNN), residual network (ResNet) and temporal attention mechanisms (TA). The GCNN is employed to capture the temporal trend features and local detail features from the fused time series data processed by the developed temporal feature fusion method. To enhance the temporal modelling capability, the ResNet is incorporated into the BiLSTM network to represent the relationship between operational parameters and unthreshing rate. Furthermore, a TA is adopted to adaptively focus on the critical time steps based on the evolutionary trends of the unthreshing rate. Ablation tests demonstrate the effectiveness of the aforementioned modules in improving the prediction performance. Compared to several benchmark models, the proposed BiLSTM-based model has better prediction performance with the R 2 , RMSE and MAE reaching 0.931±0.006, 0.053±0.012 and 0.051±0.010, respectively. Interpretability analysis confirms the model effectively captures the nonlinear coupling between operational parameters and wheat unthreshing rate. The multi-condition bench tests further confirm the effectiveness of the proposed model with all mean R 2 values remaining at or above 0.841. This study provides a theoretical and technical foundation for loss suppression control in wheat combine harvesting.