轴承故障稀疏编码特征提取与多分类SVM识别

Spare Auto-Encoder Feature Extraction and Multi-Classification SVM Recognition of Bearing Fault

为了准确识别滚动轴承故障状态,提出了基于稀疏编码器的自动特征提取方法和基于投票法多分类孪生支持向量机的故障类型识别方法。稀疏自动编码器通过对输入信号编码过程,自动学习隐藏在输入信号中的特征量,无需任何先验知识和专家经验。将投票法与孪生支持向量机相结合,提出了投票法多分类孪生支持向量机的故障模式识别方法,既发挥了投票法"民主决策精度高"的优势,同时具有孪生支持向量机训练速度快的优点。挑选了凯斯西储大学在10类故障状态下的实验数据进行验证,投票法多分类孪生支持向量机故障识别精度为99.40%,而使用神经网络故障识别精度为95.61%,比多分类孪生支持向量机降低了3.96%;投票法多分类孪生支持向量机训练时间为34.79s,而神经网络训练时间为89.76s,是多分类支持向量机的2倍以上。实验证明了投票法多分类支持向量机具有极高的故障识别精度和较少的训练时间。

To recognize bearing fault state accurately,feature auto extraction method based on spare encoder and fault type recognition based on voting multi-classification twin SVM are proposed.Through coding process of input signal,spare encoder can learn feature hiding in input signal,and it does not need any priori knowledge and expertise.Combining voting method and twin support vector machine,fault type recognition method based on voting multi-classification twin SVM is put forward.The method balances the democratic decision-making and high accuracy advantage of voting method with fast training speed of twin SVM.Case Western Reserve University experimental data under 10 type fault are selected to clarify the fault recognition method.Recognition accuracy of voting multi-classification twin SVM is 99.40%,and it is 95.61%by neutral network,which means recognition accuracy of voting multi-classification twin SVM increase by 3.96%compared with neutral network.Training time-cost of voting multi-classification twin SVM is 34.79 s,and it is 89.76 s by neutral network,which means training time-cost of voting multi-classification twin SVM is less half than neutral network.The experiment proves that voting multi-classification twin SVM posses high fault recognition accuracy and less training time-cost.