Ten acoustic emission characteristic parameters were extracted for crack monitoring by means of BP neural network of three layers. By comparing the training error and the training times of BP neural network, it was found that when the number of hidden neurons is 13, the approaching effect of BP neural network is the best and the network error is the smallest. According to the crack sensitivity of the signal, that the AE parameters represent, the AE parameters were gradually eliminated through selection to reduce the dimensions of input signal. Finally, the energy rate, mean signal level, amplitude, relative arriving time, time duration and rising count were selected as acoustic emission characteristic parameters to identify the crack of deep drawing parts. The research has theoretical and practical significance to the online monitoring of cracks in the drawing process.%运用设计的三层BP神经网络对采集到的10个声发射参数进行特征提取.通过对比不同隐含层神经元个数的BP神经网络的训练误差与训练次数,确定当隐含层神经元个数为13个时,BP神经网络的逼近效果较好,产生的网络误差最小.再利用计算各声发射参数对表征裂纹信号灵敏度的大小,逐步删除各个声发射参数,降低模式识别时输入信号的维数.最后确定相对到达时间、幅度、能率、上升计数、持续时间和平均信号电平六个声发射参数能够有效地识别金属拉深件裂纹.该研究对于金属拉深件裂纹的在线监测具有理论和实际意义.
展开▼
