基于特征迁移学习的提升机轴承智能故障诊断

Intelligent fault diagnosis of hoist bearing based on feature transfer learning

  • 摘要: 针对提升机复杂实际工况导致的现有故障诊断方法准确率低和适应性弱的问题,提出了一种基于深度迁移特征选取(DTF)与平衡分布自适应(BDA)的提升机轴承智能故障诊断方法。对不同工况下的轴承故障信号进行时频分析,提取时域、频域统计特征,采用深度置信网络进行高维深度特征提取。为从高维深度特征集中选取出既有利于故障模式识别,也有利于跨域故障诊断的特征,采用基于ReliefF与域间差异的迁移特征选取(TFRD)方法对各特征的可迁移性进行量化评估,利用TFRD方法对各特征进行类别区分度和域不变性量化评估,采用ReliefF算法处理各类特征数据,获得表征类别区分度的权重值;计算同一特征在不同域间的最大均值差异,构建一种新的特征可迁移性量化指标。基于TFRD 方法,选取特征可迁移性大的深度特征构建特征子集,利用BDA对源域和目标域的特征子集进行分布适应,降低两者间的分布差异。采用源域特征集训练故障模式识别分类器,对目标域样本进行故障识别与分类。采用经典机器学习方法、深度学习方法和迁移学习方法构建了8种故障诊断模型,用于与提出的DTF−BDA故障诊断模型进行故障诊断准确率对比。结果表明:① DTF−BDA故障诊断模型能够取得明显优于其他对比模型的性能,最高故障诊断准确率可达100%。② TFRD方法能有效提高基于迁移学习方法构建的故障诊断模型的性能,与迁移成分分析和联合分布自适应相结合情况下的最高故障诊断准确率分别可达96.46%和97.67%。

     

    Abstract: The complex actual working conditions of the hoist causes the problems of low accuracy and weak adaptability of existing fault diagnosis methods. In order to solve these problems, an intelligent fault diagnosis method of hoist bearing based on deep transferable feature selection(DTF) and balance distribution adaptation(BDA) is proposed. The bearing fault signals under different working conditions are subjected to time-frequency analysis. The time and frequency domain's statistical characteristics are extracted. The high-dimensional depth characteristics are extracted by adopting a deep belief network. In order to select features that are beneficial to fault mode identification and cross-domain fault diagnosis from a high-dimensional depth feature set, the transferable feature selection based on ReliefF and differences between domains(TFRD) method is adopted. The method carries out the quantitative evaluation of the transitivity of each feature. The TFRD method carries out the quantitative evaluation on the class discrimination and domain invariance of each feature. The ReliefF algorithm processes various feature data to obtain weight values representing class discrimination. This method calculates the maximum mean discrepancy of the same feature between different domains, and constructs a new quantitative index of feature transferability. Based on the TFRD method, depth features with high feature transferability are selected to construct feature subsets. The balance distribution adaptation is applied to carry out distribution adaptation on the feature subsets of the source domain and the target domain, so as to reduce the distribution difference between the two domains. The source domain feature set is used to train the fault pattern identification classifier, and the target domain samples are used for fault identification and classification. Eight fault diagnosis models are constructed by using the classical machine learning method, deep learning method and transfer learning method. The models are used for comparing the fault diagnosis accuracy with the proposed DTF-BDA fault diagnosis model. The results show the following points. ① The DTF-BDA fault diagnosis model can achieve better performance than other models, and the highest fault diagnosis accuracy can reach 100%. ② The TFRD method can effectively improve the performance of the fault diagnosis model based on the transfer learning method. The highest fault diagnosis accuracy can reach 96.46% and 97.67% respectively when combined with the transfer component analysis and joint distribution adaptation.

     

/

返回文章
返回