论文部分内容阅读
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300~400kHz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。
Aiming at the problems of low prediction accuracy and large error in the traditional BP (Back Propagation) neural network prediction model for grinding surface roughness, the RMS value, FFT value, standard deviation, variance and skewness of the acoustic emission signal The parameter is input unit, a three-layer BP neural network is established to predict the grinding surface roughness, and the method is improved by using additional momentum method and adaptive learning rate method. The number of hidden layer units was optimized by simulation, and the acoustic emission signal samples of 10 frequency bands were optimized by using the model. The acoustic emission signal of 300-400 kHz was determined as the learning sample frequency band of surface roughness prediction model . The experimental results show that compared with the traditional BP model, the improved BP prediction model has the characteristics of fast convergence and high prediction accuracy, and the relative error can be controlled within 8.66%.