基于卷积神经网络的超声金属材料辨识

Ultrasonic metal material identification based on convolutional neural network

传统的金属材料辨识方法会给被检测样品带来一定程度的损伤。文章通过采集金属材料的超声回波时域信号,采用短时傅里叶变换对其进行时频分析,得到包含金属材料微观组织信息的超声时频谱。将目标样品的超声时频谱预处理后作为训练样本,输入到构建好的卷积神经网络中进行训练。然后采集目标样品以及干扰样品的超声频谱图,分别将其输入网络进行辨识。结果表明,神经网络在训练时收敛较快,损失函数在迭代200次后收敛,在经过100次迭代后训练集准确率趋于100%。训练完成的网络模型记录着对应训练样本的特征信息,利用该训练好的网络对待测样本进行辨识,最终可实现超声金属材料辨识。

Traditional identification methods of metal materials will cause a certain degree of damage of the tested samples.In this paper,the ultrasonic echo signals of metal materials is collected,the short-time Fourier transform(STFT)is used for time-frequency analysis of the signals,and the ultrasonic time-frequency spectrums containing the microstructure information of metal materials are obtained.The images of the ultrasonic spectrum are processed as samples to input into the constructed convolutional neural network for training.The results show that the neural network converges quickly and can show excellent performance and recognition accuracy in the verification set.The network models after training all record the characteristic information of corresponding training samples,which can be stored as a database.Ultrasonic metal material identification can be realized by using the trained network to identify test samples.