用声发射信号和改进的BP神经网络预测磨削表面粗糙度

Prediction of Grinding Surface Roughness with Acoustic Emission Signal and Modified BP Neural Network

针对磨削表面粗糙度传统BP（Back Propagation）神经网络模型在线预测时存在预测精度低、误差大等问题,以磨削声发射信号的RMS值、FFT值、标准差、方差和偏斜度5参量为输入单元,建立了三层BP神经网络来预测磨削表面粗糙度,并应用附加动量法和自适应学习速率法对其进行了改进。通过仿真优化了隐层单元数,利用模型对磨削加工10个频段的声发射信号样本进行优选,确定将300400kHz的声发射（Acoustic Emission,AE）信号作为表面粗糙度预测模型学习样本频段。实验结果显示：改进后的BP预测模型与传统BP模型相比,具有收敛速度快、预测精度高的特点,相对误差可控制在8.66%以内。

The traditional BP model for online predicting grinding surface roughness has some shortcomings such as low prediction accuracy,big error.Three-layer BP neural network is established through RMS,FFT peak,standard deviation,variance and skew of acoustic emission signal.Methods of additional momentum and self-adapting learning rate are employed to modify the model.Amount of hidden layer unit is optimized through emulation.AE signal between 300 kHz and 400 kHz is selected as learning patterns frequency through optimization among 10 frequencies by the model.Results show that modified BP model has the merits of faster convergence rate and higher forecast precision than that of traditional BP model.And its relative error is less than 8.66%.