| 文娟.基于改进分类器动态选择算法的滚珠丝杠副状态识别[J].高技术通讯(中文),2024,34(4):396~405 |
| 基于改进分类器动态选择算法的滚珠丝杠副状态识别 |
| Ball screw condition recognition based on an improved dynamic classifier selection method |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 04. 007 |
| 中文关键词: 分类器动态选择； 邻域成分分析(NCA)； 状态识别； 滚珠丝杠副； 多分类器系统 |
| 英文关键词: dynamic classifier selection, neighborhood components analysis(NCA), condition recognition, ball screw, multiple classifier system |
| 基金项目: |
| 作者 | 单位 | | 文娟 | (浙江工业大学机械工程学院杭州 310023)
(恒丰泰精密机械股份有限公司温州 325000) |
|
| 摘要点击次数: 478 |
| 全文下载次数: 413 |
| 中文摘要: |
| 为提升滚珠丝杠副的性能状态识别精度，提出一种改进的分类器动态选择算法。该算法借助邻域成分分析(NCA)，准确并自适应地定义测试样本的邻域，无需选择距离度量方式，从而更加准确地衡量多分类器系统中各子分类器对于测试样本进行正确分类的潜力，解决了传统分类器动态选择算法精度受限于距离度量方式选择是否合适的问题。将所提出的分类器动态选择算法应用于滚珠丝杠副状态识别中，首先利用AdaBoost算法离线训练反向传播（BP）神经网络集合，然后依据实时信号特征，采用改进的分类器动态选择算法从分类器集合中选取最合适的子分类器进行状态鉴定，从而实现更好的识别效果。实验结果表明，提出方法的状态识别准确率能够达到97.22%，高于BP神经网络、AdaBoost与传统分类器动态选择算法，且对于不同的性能状态均有较高的识别精度。 |
| 英文摘要: |
| To improve the condition recognition rates of the ball screw,an improved dynamic classifier selection method is proposed. In this method, with neighborhood components analysis(NCA), the neighborhood of the test sample is defined accurately and adaptively without selecting the distance metric, and then the competence of each classifier in the multiple classifier system for accurately recognizing the testing task can be measured more exactly. Consequently, the classification accuracy is no longer restricted by the distance metric selection. The presented approach is applied to identify the health state of the ball screw. First, the AdaBoost algorithm is employed to create a back propagation (BP) neural networks pool. Then, to enhance the classification rates, the proposed dynamic classifier selection methodology is utilized to select the most suited classifier from the classifier pool for condition recognition according to the features extracted from the online signal. Experimental results show that the proposed method can identify the ball screw condition effectively with an accuracy of 97.22%, which is higher than that of the BP neural networks, AdaBoost, and the conventional dynamic classifier selection method. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
