A Prediction Scheme for Loader Excavation Resistance Based on Parallel Hybrid Deep Neural Network

In response to the challenge of directly measuring excavation resistance, this study introduces a predictive method based on a parallel hybrid deep neural network. Initially, addressing the difficulty in obtaining excavation resistance data, a joint electromechanical-hydraulic simulation model for loaders was developed to generate a dataset for excavation resistance. Considering the presence of local spatial features and long-term dependencies in the sequence, a residual structure was incorporated into the convolutional neural network (CNN) to enhance the focus on significant features in the resistance data during spatial feature extraction. Additionally, gated recurrent units (GRUs) were employed to capture temporal features. Subsequently, these two types of features were separately unfolded and then integrated through a fully connected layer to facilitate feature fusion, thereby avoiding the degradation of original data features typical in traditional serial architectures. Finally, the predictive performance of the proposed method was evaluated in comparative experiments with CNNs, GRUs, Transformers, and serial-structure hybrid neural networks. The results demonstrated an improvement in prediction accuracy compared to other methods. Specifically, the weighted mean absolute percentage error (WMAPE) values for longitudinal resistance (Fx), vertical resistance (Fz), and resistance torque (My) were 2.89%, 3.5%, and 5.78%, respectively. The proposed method decreases the WMAPE by 45.47%, 44.35%, and 26.8% in comparison to the serial architecture. The mean ${R}^{{2}}$ values were 0.982, 0.977, and 0.962, respectively.