Activation functions selection for BP neural network model of ground surface roughness
Activation functions selection for BP neural network model of ground surface roughness
复制标题
地表粗糙度BP神经网络模型激活函数选择
DOI：
10.1007/s10845-020-01538-5
复制
发表时间：
2020-01-30
影响因子：
8.3
通讯作者：
Guo, Dongming
中科院分区：
文献类型：
--
作者：
Pan, Yuhang;Wang, Yonghao;Guo, Dongming
磨削表面粗糙度预测是了解和优化磨削加工过程的关键。然而，由于磨削过程的复杂性，理论模型和经验模型很难准确地预测磨削表面粗糙度。BP神经网络可用于建立加工参数与表面粗糙度之间的关系，避免了揭示复杂物理机理的困难，因此在工业实践中磨削工艺的自动优化具有独特的潜力。激活函数是影响BP神经网络效率和精度的重要因素之一。然而，它通常是任意选择的，或者最多通过试验或调整来选择。本文提出了一种激活函数选择方法，利用由近似物理模型产生的虚拟数据来评估BP神经网络在实际应用中的性能。结果表明，以tansig作为隐含层的激活函数，以purelin作为输出层的激活函数，BP神经网络模型可以获得最高的学习效率。此外，当隐含层激活函数为形状因子为1~3的S型函数，而输出层激活函数为纯净时，模型的预测精度较高。最后，通过比较虚拟数据和实验数据得到的BP神经网络的性能，验证了该方法的有效性。结果表明，该方法是确定BPNN激活函数的一种简单有效的方法。
Roughness prediction of ground surfaces is critical in understanding and optimizing the grinding process. However, it is hitherto difficult to predict accurately the ground surface roughness by theoretical and empirical models due to the complexity of grinding process. BP neural network (BPNN), which can be used to establish the relationship between processing parameters and surface roughness, avoids the difficulty of revealing the complex physical mechanism and thus has unique potential in automatic optimization of grinding process in industrial practice. Activation function is one of the most important factors affecting the efficiency and accuracy of BPNN. Nevertheless, it is often selected arbitrarily or at most by trials or tuning. This paper proposes an activation function selection approach in which virtual data generated from the approximate physical model are employed to evaluate the performance of the BPNN in practice application. The results show that with tansig as the activation function of hidden layer and purelin as the activation function of output layer, the BPNN model can obtain the highest learning efficiency. Moreover, when the activation function of hidden layer is sigmoid, whose shape factor is 1-3, and the output layer activation function is purelin, the model can predict more precisely. Finally, the proposed approach is validated by comparing the performance of BPNN obtained from the virtual data and the experimental data. Obtained results showed that the proposed approach is a simple and effective way to determine the activation function of BPNN.