基于ResNet多特征图融合的钻削表面粗糙度分类方法

ResNet-based multi-feature map fusion for classification of drilling surface roughness

传统五面复合数控(CNC)钻削表面粗糙度测量工作复杂,采用人工测量存在较大人为误差。传统多元回归、多项式拟合方法仅采用转速和进给速度参数,数据利用率低且噪声敏感性强;用传统机器学习方法无法有效提取信号的深层复杂特征。针对上述问题,提出了一种基于ResNet模型、频谱图特征与时频图特征融合的钻削表面粗糙度分类预测方法。首先,根据CNC钻削加工理论和企业实际CNC钻削经验确定了CNC钻削加工实验的工艺参数变量;然后,基于SYNTEC CNC系统开发了多源数据采集系统,实时采集了钻削加工过程数据;接着,分析了三轴振动信号的频谱特征和时频特征,验证了振动信号跟表面粗糙度类别的关联性;随后,采用卡尔曼滤波对三轴振动信号进行了降噪处理,采用快速傅里叶变换(FFT)和连续小波变换(CWT)进行了振动信号频谱热图与时频图转换,采用矩阵拼接对三轴振动信号的单轴时频图进行了拼接融合,得到了三轴振动时频图;最后,对频谱热图和时频图进行了卷积运算融合频谱特征与时频特征,并进行了ResNet和其他网络模型如Densenet、Shufflenet和Mobilenet_v3_small等的对比实验。研究结果表明:相对上述其他网络模型,基于ResNet网络模型的表面粗糙度分类正确率提高了约9%,同时也验证了三轴时频特征融合以及频谱特征和时频特征融合方法的正确性。由于模型训练成本低、训练收敛速度快,该方法在轻量级、低成本的CNC机床钻削表面粗糙度预测分类中具有良好的工业应用前景。

The traditional five-face composite computerized numerical control(CNC) drilling surface roughness measurement is complicated,and there is a large human error in manual measurement.The traditional multiple regression and polynomial fitting methods only use rotational speed and feed speed parameters with low data utilization and high noise sensitivity;traditional machine learning can not effectively extract the deep and complex features of the signal.Aiming at the above problems,a classification and prediction method of drilling surface roughness based on ResNet model,fusion of spectrogram features and time-frequency graph features was proposed.Firstly,the process parameter variables of the CNC drilling processing experiment were determined according to the theory of CNC drilling processing and the actual CNC drilling experience of the enterprise.Secondly,a multi-source data acquisition system was developed based on SYNTEC CNC system,and the drilling process data were collected in real time.Then,the spectral and time-frequency characteristics of the three-axis vibration signals were analyzed,and the correlation between the vibration signals and the surface roughness category was verified.Then,the Kalman filtering was used for noise reduction of the three-axis vibration signals,and the fast Fourier transform(FFT) and the continuous wavelet transform(CWT) were used to convert the spectro-thermograms and time-frequency maps of the vibration signals,and matrix splicing was used to splice and merge the uniaxial time-frequency maps of the three-axis vibration signals to get the three-axis vibration time-frequency map.Finally,the fusion of spectral and time-frequency features was realized by convolving the spectral heat map and time-frequency map,and the comparison experiments between ResNet and other network models such as Densenet,Shufflenet and Mobilenet_v3_small were carried out.The research results show that the correct rate of surface roughness classification based on the ResNet network model is improved by about 9% relative to the other network models mentioned above,and the correctness of the three-axis time-frequency feature fusion as well as the fusion method of spectral and time-frequency features is also verified.Due to the low cost of model training and fast training convergence,the method has a good prospect for industrial application in lightweight and low-cost prediction and classification of surface roughness of drilling on CNC machine tools.