基于流形特征域适配的滚动轴承故障诊断

Rolling bearing fault diagnosis based on manifold feature domain adaptation

针对变工况下存在两域特征分布复杂,数据在原始空间进行分布对齐时,特征扭曲和发散难以消除等问题,提出一种基于流形特征域适配(manifold feature domain adaptation, MFDA)的滚动轴承故障诊断方法。该方法通过无监督的方式生成一个与目标域具有相似分布的中间域,构建与源域、中间域和目标域相关的公共子空间,并利用局部生成差异度量保留数据在子空间中的流形局部几何结构,以避免数据对齐时出现扭曲和发散;同时利用最大均值差异度量对齐中间域和目标域,以最小化两域间的分布差异,保证数据间局部与全局结构的相关性。最后,利用学习到的特征,以最小二乘法实现滚动轴承的跨域故障识别。在三组滚动轴承数据集上进行试验验证,与其他智能识别算法相比,该方法能有效避免特征扭曲和发散,且具有优良的泛化性能。

Here,aiming at complex distribution of two domain features under variable working conditions and difficulty in eliminating feature distortion and divergence when aligning data in original space,a rolling bearing fault diagnosis method based on manifold feature domain adaptation(MFDA)was proposed.With this method,an intermediate domain with a similar distribution to target domain was generated in an unsupervised manner,a common subspace related to source domain,intermediate domain and target domain was constructed,and the manifold local geometric structure of data in the subspace was reserved using local generation difference measurement to avoid distortion and divergence during data alignment.Simultaneously,the maximum mean difference measure was used to align intermediate domain and target domain,minimize distribution differences between the two domains and ensure correlation between local and global structures of data.Finally,using learned features,cross-domain fault identification of rolling bearings was realized with the least squares method.Experimental verification was performed on 3 rolling bearing datasets.It was shown that compared with other intelligent recognition algorithms,the proposed method can effectively avoid feature distortion and divergence,and have excellent generalization performance.