中文摘要：

信号的时域统计特征是最早应用且最为简洁实用的特征参数。目前,用于模式识别的时域统计特征的选用多是基于经验或者不加选择地使用,识别正确率难以保证。针对这个问题,提出一种可视化的特征优选方法。该方法根据特征数据的轮廓图,分析各维特征数据的聚类特性,去除聚类性弱、对故障区分无益的冗余特征维度,仅保留聚类性强的特征维度用于故障识别。在轴承故障诊断实验中,对故障信号时域统计特征进行优选,并采用BP神经网络进行故障识别。实验结果表明,优选后故障识别率得到大幅度提高。

英文摘要：

Time domain statistical features are the earliest used and the most simple and practical feature parameters.At present, time domain features for pattern recognition are chosen on experience or used by no choice.As the result, the fault recognition rate could not be ensured.Aiming at this problem, it proposes a method of visual feature selection, which analyzes the clustering characteristics of the feature data according to the contours of the data of various classes of features and removes the redundant feature dimensions.The method also retains the feature dimensions with strong clustering characteristics for the fault recognition.In the bearing fault diagnosis experiment, the features of time-domain statistics of fault signals are optimally selected, and BP neural network is used for fault pattern recognition.The fault recognition rate is significantly higher than the fault diagnosis without selection of features.The experimental results show that the proposed method is simple and feasible, and possessed of popularization and application value.