基于时频分析和深度学习的表面粗糙度超声模式识别方法

A Deep Learning-Based Electromagnetic Ultrasonic Recognition Method for Surface Roughness of Workpeice

表面粗糙度直接决定着工件的性能和使用寿命,由于传统的基于光学或三维形貌的表面粗糙度检测方法存在对工件表面清洁状态及操作环境要求较高等问题,因此,该文提出一种基于深度学习的非接触式电磁超声表面粗糙度识别方法。首先通过建立不同表面粗糙度的电磁超声有限元仿真模型,研究了涡流密度和洛仑兹力对激励与接收信号的影响。然后利用所提出的卷积神经网络,对从电磁超声换能器检测得到的A扫描信号的时频系数图进行特征提取,输入至预训练的支持向量机分类器中完成表面粗糙度识别和预测。为了验证所提方法的有效性,对通过立铣工艺加工的表面粗糙度比较样块进行测试。实验结果表明,所提出的超声识别方法平均精度为98.83%,具有较高的预测精度与稳定性,解决了超声信号信噪比较低而导致信号特征识别困难的问题,同时减少了特征提取过程对于人工干预的依赖。

Surface roughness directly determines the performance and service life of the workpiece.As customary surface roughness detection methods based on optical or three-dimensional profilometers have higher requirements on the surface cleaning state and operating environment, a non-contact electromagnetic ultrasonic surface roughness recognition method based on deep learning is proposed under this paper. Firstly, the effects of eddy current density and the Lorentz force on the excitation and reception signals are investigated by establishing the finite element simulation model of electromagnetic ultrasound with different surface roughness. Then, the proposed convolutional neural network is utilized to extract the features of the time-frequency coefficient map of the A-scan signal detected by the electromagnetic ultrasonic transducer, which is input into the pre-trained support vector machine classifier to complete the roughness recognition and prediction. To verify the proposed method, the surface roughness comparison block processed by the end milling process is tested. The experimental results show that the average accuracy of the proposed ultrasonic recognition method is 98.83%, which has high prediction accuracy and stability, solves the problem of the low signal-to-noise ratio of the ultrasonic signal which leads to difficult signal feature recognition, and reduces the dependence of feature extraction process on manual intervention.