Authors

Abstract

The current field of pneumatic subventing prediction focuses on a single task and neglects the possible interrelationships between different outputs. In order to improve the prediction accuracy and reduce the number of algorithm model establishment, this study conducted field experiments on soil in autumn and winter. Neural network algorithms RBF (radial basis neural network), BP (backward propagation neural network), DNN (Deep learning network) and CNN (Convolutional neural network) were used to make multi-output regression prediction for changing the traction resistance and disturbance area affected by different levels of subsooning velocity, depth and pressure value in the process of pneumatic subsooning. The evaluation indexes RMSE, MAE and R2 were compared with the single output regression model, and the accuracy of the four models with the highest accuracy was compared with that of its own single output model to prove the correlation between traction resistance and disturbance area. The results showed that the R2 of the four model test sets of RBF, BP, DNN and CNN were 0.9999, 0.9966, 0.9986 and 0.9762, respectively. The R2 of the disturbance area are 0.9997, 0.9924, 0.9968 and 0.9715, respectively. RBF has the highest R2 and the lowest RMSE and MAE, indicating that the RBF model has the best prediction effect. Compared with the single output regression model of RBF model, the prediction accuracy of both outputs is higher, so it can be used to predict the subsoiling drag resistance and disturbance area.

Abstract in Chinese

针对目前气动深松预测领域多聚焦于单一任务，忽略了不同输出之间可能存在的相互关系。为了提高预测的精确性，并且减少算法模型建立次数，本研究对秋冬两个季节下的土壤进行田间试验，利用神经网络算法RBF（径向基神经网络）、BP（逆向传播神经网络）、DNN（深度学习网络）、CNN（卷积神经网络）对气动深松过程中改变受不同水平的深松速度、深度、气压值影响的牵引阻力及扰动面积值进行多输出回归预测，利用评价指标RMSE、MAE、R2与单输出回归模型进行对比评估，将四个模型中精度最高的与本身的单输出模型的精度进行对比，证明牵引阻力及扰动面积之间的相关性。结果表明：RBF、BP、DNN、CNN四个模型测试集牵引阻力的R2分别为0.9999、0.9966、0.9986、0.9762。扰动面积的R2分别为0.9997、0.9924、0.9968、0.9715。RBF的R2最高，RMSE、MAE最低，可见RBF模型预测效果效果最好，且相较于RBF模型的单输出回归模型两个输出的预测精度都较高，因此可用于深松牵引阻力及扰动面积的预测。