›› 2021, Vol. 27 ›› Issue (6): 1629-1640.DOI: 10.13196/j.cims.2021.06.009

Previous Articles     Next Articles

Crankshaft intelligent recognition method based on deep support vector machine

  

  • Online:2021-06-30 Published:2021-06-30
  • Supported by:
    Project supported by the Hebei Provincial Natural Science Foundation,China(No.E2017202294),the Hebei Provincial Youth Top Talent Foundation,China(No.210014),and the National Natural Science Foundation,China(No.51305124).

基于深度支持向量机的曲轴智能识别方法

杨泽青1,王春方1,彭凯1+,刘丽冰1,张亚彬2   

  1. 1.河北工业大学机械工程学院
    2.机械工业仪器仪表综合技术经济研究所
  • 基金资助:
    河北省自然科学基金资助项目(E2017202294);河北省青年拔尖人才资助项目(210014);国家自然科学基金资助项目(51305124)。

Abstract: In view of recognizing crankshaft types with similar topological shapes and structures and only local differences in details,Deep Support Vector Machine (DSVM) recognition method based on feature fusion was proposed.The method combined deep neural networks with multiple SVMs to form a network model,which extracted deep features by maximizing the use of support vector structure risk minimization principles to establish complex nonlinear mapping relationships between features and target values to ensure the generalization ability of the model.This model contained the data input layer,hidden layer and output layer.To obtain a better edge detection of local details in crankshaft images,the traditional Canny edge detection algorithm was improved from filtering,gradient calculation,automatic acquisition of high and low thresholds,and then the Hu moment,Fourier descriptor and size features were extracted.The input vector of DSVM model was obtained by fusing and normalizing the extracted features for training the shallowest SVM.The upper level features were generated from the support vector mapping at the lower level to achieve layer-by-layer training learning,and the network was updated by the backpropagation algorithm,thus the classification and recognition results of the network were output by discriminant function.In addition,the sample data sets of five crankshafts were established by properly designing the crankshaft image acquisition scheme,and the performance of the model was verified.The experimental results showed that the recognition accuracy of the model could reach 99.6%,which was 6.6% and 3.1% higher than the single SVM and AlexNet respectively,and the recognition time was 93ms.It met the requirements of crankshaft identification in the flexible production line of remanufacturing waste parts repair or transformation.

Key words: deep support vector machine, intelligent recognition, crankshaft, feature fusion, remanufacturing

摘要: 针对拓扑形状结构相似、仅在局部细节上有差别的曲轴识别问题,提出一种基于特征融合的深度支持向量机(DSVM)识别方法,该方法将深度神经网络与多个支持向量机(SVM)相结合构成一种网络模型,通过最大限度地利用支持向量结构风险最小化原理提取深层特征,以建立特征和目标值之间的复杂非线性映射关系,保证模型的泛化能力。该模型包含数据的输入层、隐藏层和输出层,为获得较好的曲轴图像局部细节边缘检测效果,从滤波、梯度计算、自动获取高低阈值等方面对传统Canny边缘检测算法进行改进,进而提取边缘的Hu矩、傅里叶描述子和尺寸特征,通过串行融合和特征筛选方法对提取到的3类特征进行优化组合并做归一化处理,作为DSVM模型的输入向量用于训练最浅层的SVM;高层的特征由低层的支持向量映射产生,实现逐层的训练学习,通过反向传播算法对网络进行更新,由判别函数输出网络的分类识别结果。通过设计曲轴图像采集方案,建立了5类曲轴样本数据集,并验证了模型的性能。实验结果表明,该模型识别精度可达99.6%,相较于单一的SVM和AlexNet分别提高了6.6%和3.1%,识别时间为93 ms,符合再制造废旧零件修复或改造柔性生产线中对曲轴识别的要求。

关键词: 深度支持向量机, 智能识别, 曲轴, 特征融合, 再制造

CLC Number: