视听信息融合的CNC铣刀磨损状态监测方法

A CNC tool wear monitoring method based on audio-visual information fusion

针对目前多特征融合刀具磨损监测方法中存在特征之间相关性差、非线性关系被忽略,导致用于模式识别的融合特征维数过大、冗余信息多、特征契合度差和识别准确率低的问题。提出了一种核典型相关分析(Kernel Canonical Correlation Analysis,KCCA)的多级特征融合方法。采集数控加工过程的声压信号以及工件纹理图像,并提取相应的视听特征,利用核典型相关分析法在高维空间找到2组投影方向,保证投影后特征间的皮尔逊系数最大,使视听特征的相关性最大化。经实验验证,利用核典型相关分析法能计算6组典型变量,并表达原特征97%以上的信息,大大降低了特征维数、减少了冗余特征。并且同样的识别模型下,核典型相关分析法能够将检测准确率提升至95%以上。

In the current multi-feature fusion tool monitoring,the correlation between features is poor and the nonlinear relationship is ignored,which leads to the problems of excessive fusion feature dimension,excessive redundant information,poor feature fit and low recognition accuracy.A multi-level feature fusion method for Kernel Canonical Correlation Analysis(KCCA)was proposed.The sound pressure signal and workpiece texture image of CNC machining process were collected,and the corresponding audio-visual features were extracted,and the kernel canonical correlation analysis method was used to find two groups of projection directions in the high-dimensional space,so as to ensure the maximum Pearson coefficient between the projected features and maximize the correlation of audio-visual features.The experimental results show that kernel canonical correlation analysis method can express more than 97%of the information of the original features with 6 groups of typical variables,which greatly reduces the feature dimension and redundant features.Moreover,under the same recognition model,kernel canonical correlation analysis method can improve the detection accuracy to more than 95%.