柴油机故障的堆栈自编码特征提取与随机森林识别

Stack Auto Encoding Feature Extraction and Random Forest Recognition of Diesel Fault

为了提高柴油机故障的在线识别准确率,提出了堆栈自编码特征提取方法和话语权随机森林的故障类别识别方法。对缸盖罩振动信号为分析对象,提出了样本熵自适应小波阈值去噪方法,有效提高了信号的信噪比。使用堆栈自编码网络提取振动信号的故障特征向量,所提取特征类内聚合度高、类间区分度好。在传统森林算法基础上,根据决策树的预测试准确率为其赋予不同的话语权,从而提出了话语权随机森林算法,并将其应用于柴油机运行故障模式识别。经10组实验验证,传统森林算法的平均识别准确率为90.32%,话语权森林算法的平均识别准确率为99.67%,比传统算法提高了10.35%;另外,话语权森林算法的识别准确率标准差远小于传统随机森林算法。以上数据说明经过改进,随机森林算法的识别准确率和稳定性均得到了提高。

In order to improve online recognition accuracy of diesel fault,feature extraction method of stack auto encoding and fault model recognition based on speaking-weight forest algorithm are proposed.Selecting rocket cover vibration signal as analyt⁃ic target,sample entropy adaptive wavelet threshold de-noising method is put forward,which improves SNR validly.Stack auto encoding network is used to extract fault feature of vibration signal,and polymerization degree in the category is very high,dis⁃crimination between the category is obvious.On the basis of traditional random forest algorithm,different speaking-weight is giv⁃en to every decision tree relying on forecasting accuracy,so that speaking-weight forest algorithm is proposed,and the algorithm is used to recognize diesel fault model.10 group experiment is executed,even recognition accuracy of traditional random forest al⁃gorithm is 90.23%,and even recognition accuracy of speaking-weight forest algorithm is 99.67%,it is 10.35%more than the val⁃ue of traditional random forest algorithm.Besides,recognition accuracy standard deviation of speaking-weight is much less than traditional random forest.The data above indicates that recognition accuracy and stability of improved random forest algorithm is enhanced.