基于支持向量机的颤振在线智能检测

发布时间：2018-08-25 10:45

【摘要】：为了检测车削过程中的颤振,提出一种颤振在线智能检测方法。使用最小二乘一类支持向量机,训练出描述特征矢量集的超球面,通过计算被测样本与超球面的距离来判断其是否颤振。基于相干准则和分块矩阵求逆,构造了在线稀疏结构的最小二乘一类支持向量机,将特征信息存储于特征库(字典)中,通过更新特征库实现检测模型的在线进化。在颤振检测的应用中,首先使用小波包分解,得到第三层节点能量的比例作为特征矢量,以离线数据构造特征矢量作为输入,训练得到初始检测模型以及特征库,在线检测中不断更新特征库,实现检测模型的在线进化。试验结果表明,在车削颤振识别中,在线进化的检测模型的识别效果更好,颤振预报准确率高达至99.04%,优于离线模型的预报准确率96.74%。
[Abstract]:In order to detect chatter in turning process, a method of on-line intelligent flutter detection is proposed. A class of least square support vector machines is used to train the hypersphere to describe the set of feature vectors and to determine whether the hypersphere is flutter or not by calculating the distance between the measured sample and the hypersphere. Based on coherence criterion and block matrix inversion, a class of least squares support vector machines (LSM) with online sparse structure is constructed. The feature information is stored in the feature library (dictionary), and the online evolution of the detection model is realized by updating the signature database. In the application of flutter detection, firstly, wavelet packet decomposition is used to obtain the ratio of the energy of the third layer node as the feature vector, and the off-line data is used as the input to construct the feature vector, and the initial detection model and the signature library are obtained by training. In the online detection, the feature library is constantly updated to realize the online evolution of the detection model. The experimental results show that the on-line evolution detection model has better recognition effect in turning flutter recognition, and the accuracy of flutter prediction is up to 99.04, which is better than that of offline model.
【作者单位】：上海交通大学机械系统与振动国家重点实验室;
【基金】：国家重点基础研究发展计划资助项目(973计划,2013CB035804)
【分类号】：TG51;TP181

【参考文献】