永磁式线性力电机是直驱式电液伺服阀的重要部件之一，其驱动力可以对进入阀口的金属碎片进行切割，防止阀口被碎片挡住，因此，准确预测其驱动力对永磁式线性力电机的设计具有极其重要的研究意义。首先基于ANSOFT对永磁式线性力电机电磁场进行有限元仿真，获得其零位在极限电流作用下的驱动力。其次根据优化目标和约束条件确定永磁式线性力电机的关键结构参数及其取值范围。随后，采用基于最大最小距离准则的拉丁超立方算法进行关键结构参数在多维空间尺度上的样本采样。最后，提出带有转换层的深度神经网络模型，把电机结构参数经过转换层后提取出电机模型的100个参数，使深度神经网络能从更多的特征中组合出新的高维特征，从而提高模型预测精度，且应用PReLU激活函数和SmoothL1Loss损失函数，建立了舵机用永磁式线性力电机驱动力预测模型。与传统的预测模型Kriging和RBF相比，充分验证了此模型的有效性和准确性。

The permanent magnet linear force motor is one of the important components of the direct drive servo valve. Its driving force can cut the metal fragments entering the valve port to prevent the valve port from being blocked by them. Therefore, the accurate prediction of the driving force has extremely important research significance for designing the permanent magnet linear force motor. Firstly, a finite element simulation model of the permanent magnet linear force motor is established with ANSOFT, getting the driving force under the zero position and its limiting current. Secondly, the key structural parameters and their value ranges of the linear force motor are determined according to the optimization goal and constraint conditions. Then, the Latin hypercube algorithm based on the maximum and minimum distance criterion is used to sample data in a muti-dimensional space. Finally, a deep neural network model with a conversion layer is proposed. The conversion layer extracts 100 parameters from motor model, so that deep neural network can combine new high-dimensional features from more features and improve its prediction accuracy. The prediction model of permanent magnet linear force motor driving force used by actuator with PReLU and SmoothL1Loss is established. The comparison with traditional prediction models of both Kriging and RBF proves the effectiveness and accuracy of the new model.

国家自然科学基金联合基金项目(U20A20293）