In bearings defect diagnosis applications, information fusion has been widely used to improve identification accuracy for different types of faults, which may lead to high-dimensionality and information redundancy of the data and thus degenerate the classification performance. Therefore, it is a major challenge for machinery fault diagnosis to extract optimal features from high-dimensional and redundant data for classification. In addition, in order to guarantee the performance of fault diagnosis, conventional supervised methods usually require a large amount of labeled data available for learning. However, it is extremely difficult, costly and time-consuming to collect faulty labeled samples with class information, especially for expensive and critical machines, which often results in only a few labeled data available with a large amount of unlabeled data redundant. In this paper, we propose a novel bearing defect diagnosis model based on semi-supervised kernel local Fisher Discriminant Analysis (SSKLFDA) using pseudo labels, which can effectively extract optimal features for classification and simultaneously utilize unlabeled data for regularizing the supervised dimensionality reduction. The proposed SSKLFDA first adopts Density Peak Clustering technique to generate pseudo cluster labels for the labeled and unlabeled data and then regularizes the between-class scatter and within-class scatter according to two corresponding regularization strategies associated with the generated pseudo cluster labels. This regularization can further improve the discriminant performance of the extracted features and also make it suitable for the cases with the multimodality and noises. In order to accommodate for non-linear feature extraction, the kernel version of the proposed method is also provided with the introduction of kernel trick. The experimental results under different feature dimensions, numbers of labeled data, and subsequent classifiers scenarios demonstrate that the proposed SSKLFDA based bearings fault diagnosis model achieves higher classification performance than other existing dimensionality reduction methods-based models.

在轴承缺陷诊断应用中，信息融合已被广泛用于提高对不同类型故障的识别精度，这可能导致数据的高维和信息冗余，从而降低分类性能。因此，从高维和冗余数据中提取最佳特征进行分类是机械故障诊断的主要挑战。另外，为了保证故障诊断的性能，常规的监督方法通常需要大量可用于学习的标记数据。但是，要收集带有类别信息的错误标签样品非常困难，昂贵且耗时，特别是对于昂贵且关键的机器而言，这通常只会导致少量标记数据可用，而大量未标记数据则是多余的。在本文中，我们提出了一种基于伪监督的基于半监督核局部Fisher判别分析（SSKLFDA）的新型轴承缺陷诊断模型，该模型可以有效地提取用于分类的最佳特征，同时利用未标记的数据对监督的降维进行正则化。建议的SSKLFDA首先采用密度峰值聚类技术为标记和未标记的数据生成伪聚类标签，然后根据与生成的伪聚类标签相关的两种相应的正则化策略对类间散布和类内散布进行正则化。这种正则化可以进一步改善所提取特征的判别性能，也使其适合于具有多模态和噪声的情况。为了适应非线性特征提取，在引入内核技巧的同时还提供了所提出方法的内核版本。在不同特征尺寸，标注数据的数量以及后续分类器场景下的实验结果表明，与其他现有的基于降维方法的模型相比，基于SSKLFDA的轴承故障诊断模型具有更高的分类性能。