基于无监督迁移成分分析和深度信念网络的轴承故障诊断方法

针对轴承故障样本少导致识别精度低的问题,提出一种基于无监督迁移成分分析(unsupervised transfer component analysis,UTCA)和深度信念网络(deep belief network,DBN)的故障诊断方法。首先利用UTCA的核函数将不同工况样本特征映射到一个共享再生核Hilbert空间中,使得源域和目标域样本集更加相似,并通过最大均值偏差嵌入法(maximum mean discrepancy embedding,MMDE)判断能够迁移的源域数据,将源域样本迁移到目标域中,为深度学习提供充足的训练样本,解决了实际故障样本较少的问题;然后采用DBN模型对源域样本进行训练,再对映射后无标记的目标域样本进行故障诊断分析。利用不同工况下的滚动轴承实验数据进行算法验证,结果表明,与普通DBN、SVM、BPNN以及传统机器学习-UTCA融合方法相比,本文方法对滚动轴承故障的诊断精度更高。

Bearing fault diagnosis based on unsupervised transfer component analysis and deep belief network

To address the problem of low recognition accuracy due to insufficient fault samples, a bearing fault diagnosis method based on unsupervised transfer component analysis (UTCA) and deep belief network (DBN) is proposed. Firstly, UTCA kernel function is used to map the sample characteristics in different working conditions into a sharable reproducing kernel Hilbert space so that the sample sets in source and target domains are more similar. Then the transferable source domain data are chosen by the maximum mean deviation embedding (MMDE) method and transferred to the target domain, which provides sufficient training samples for deep learning and solves the shortage of fault samples. Finally, DBN model is applied to train the source domain samples, and the fault diagnosis is performed on the unlabeled target domain samples after mapping. The proposed algorithm is tested by the experimental data of rolling bearing in different working conditions. The results show that, compared with DBN, SVM, BPNN and those methods integrating traditional machine learning and UTCA, this algorithm has higher diagnosis accuracy for rolling bearing faults.